Awesome LLM Memory

A curated collection of papers on memory mechanisms for large language models, covering agent memory systems, retrieval-augmented generation, KV cache optimization, long-context management, and more.

Coverage: 2026-04-15 to 2026-06-04  ·  563 papers

Contents

• Agent Memory Systems (191)
• Retrieval-Augmented Generation (199)
• KV Cache & Attention Memory (56)
• Long-Context Management (21)
• Graph-Based Memory (19)
• Conversational & Personal Memory (26)
• Memory Retrieval & Search (19)
• Surveys & Benchmarks (13)
• Other Memory Systems (19)

Agent Memory Systems

Memory architectures, storage, and management for LLM-based agents handling long-horizon tasks.

• Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents 2026-06-04 conf:99% The paper proposes MAGE, a hierarchical memory system for long-horizon LLM agents that manages execution state instead of relying on semantic retrieval alone. It stores trajectories in a state tree, supports compression and revision, and improves task success while cutting token usage.

• EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents 2026-06-04 conf:99% EMBER learns what evidence a long-horizon agent should retain before future queries are known, under a fixed memory budget. It stores compact source-backed evidence capsules and improves retrieval and answer quality over budgeted baselines on a long-memory evaluation benchmark.

• TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory 2026-06-04 conf:98% TOKI formalizes contradiction resolution in LLM-agent persistent memory as write-time concurrency control. It defines bitemporal memory operators with explicit isolation and provenance guarantees, proving correctness and showing how audited memory updates can avoid replay inconsistency, belief drift, and audit loss.

• When Should Memory Stay Silent: Measuring Memory-Use Boundaries in Memory-Augmented Conversational Agents 2026-06-04 conf:98% This paper introduces RBI-Eval, a benchmark for testing when memory-augmented conversational agents should avoid using sensitive long-term memories. It shows retrieval alone is insufficient and that safe personalization requires memory-aware decisions in both retrieval and response generation.

• Beyond Similarity: Trustworthy Memory Search for Personal AI Agents 2026-06-04 conf:98% The paper shows that similarity-based long-term memory retrieval in personal AI agents can cause unsafe or inappropriate behavior. It introduces MemGate, a lightweight plug-in that filters retrieved memories based on the current task, improving safety while preserving memory utility.

• MARDoc: A Memory-Aware Refinement Agent Framework for Multimodal Long Document QA 2026-06-04 conf:98% MARDoc introduces a three-agent framework for multimodal long-document QA that replaces a single growing context with structured, dynamically updated evidence and reasoning memory. This reduces context noise and improves multi-hop reasoning accuracy over strong baselines.

• Enhancing Software Engineering Through Closed-Loop Memory Optimization 2026-06-04 conf:98% The paper proposes a closed-loop framework that evaluates and optimizes memory for LLM software-engineering agents using validated downstream impact. It improves single-episode and cross-episode performance while reducing computational cost.

• ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction 2026-06-04 conf:98% ABBEL uses recursively updated natural-language belief states as an explicit memory for sequential decision-making agents. It analyzes memory failures and improves belief updating and compression with RL-based training, narrowing the gap to full-context models while using much less memory.

• Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense 2026-06-04 conf:94% Membrane introduces a self-evolving contrastive safety memory that stores paired harmful and benign patterns for LLM guardrails. Retrieved memory cells guide safer inference without retraining, improving jailbreak defense while reducing false refusals in both standalone models and agents.

• MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery 2026-06-04 conf:93% MLEvolve is a multi-agent LLM framework for automated algorithm discovery that adds retrospective memory and dynamic global memory to reuse past experience during long-horizon search. It also improves coordination with graph-based cross-branch information flow and hierarchical planning, achieving strong results on MLE-Bench.

• LongSpace: Exploring Long-Horizon Spatial Memory from Perception to Recall in Video 2026-06-04 conf:92% Introduces LongSpace-Bench and LongSpace, a memory framework for long-video MLLMs that stores layer-aware spatial memory and retrieves it for question answering. It improves long-horizon spatial reasoning by combining chunked video processing with 3D structural cues.

• PersonaTree: Structured Lifecycle Memory for Person Understanding in LLM Agents 2026-06-03 conf:99% PersonaTree introduces a structured long-term memory for LLM agents that turns interaction evidence into stable person-level claims through a three-level persona tree. It uses conservative writing, confidence-guided consolidation, and query-conditioned retrieval to improve person understanding and persistent memory performance.

• RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation 2026-06-03 conf:99% RAMPART introduces a registry-based in-memory system for LLM agents that compiles context from addressable blocks using explicit policies for promotion, gating, writing, eviction, and rollback. It shows better task success through block placement and grouping, cuts prompt cost with relevance gating, and supports shared agent memory with zero coordination token cost.

• Scaling Self-Evolving Agents via Parametric Memory 2026-06-03 conf:99% TMEM gives LLM agents a self-evolving memory by combining explicit stored history with fast online LoRA updates that change the agent’s behavior during an episode. It frames memory extraction and adaptation as an optimizable decision process and shows gains over summary- and retrieval-based memory baselines.

• EpiEvolve: Self-Evolving Agents for Streaming Pandemic Forecasting under Regime Shifts 2026-06-03 conf:98% EpiEvolve is a self-evolving LLM forecasting agent that uses hierarchical episodic memory, delayed-label reflection, regime-aware retrieval, and strategic rule distillation to adapt without updating model weights. It improves streaming pandemic forecasting accuracy and recovers faster after regime shifts.

• From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents 2026-06-03 conf:98% This paper systematically analyzes memory poisoning in LLM agents, identifying exploitable memory write channels, architectural vulnerabilities, and attack types. It also introduces MPBench, a benchmark for evaluating these attacks, and shows that stronger memory use can increase agent exploitability.

• Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs 2026-06-03 conf:96% This paper introduces MAMA, a framework for measuring how multi-agent LLM communication topologies affect leakage of memorized private information. It shows that denser connections and shorter attacker-target paths increase leakage, and recommends sparse, hierarchical designs with topology-aware access control.

• From Agent Traces to Trust: Evidence Tracing and Execution Provenance in LLM Agents 2026-06-03 conf:84% This survey studies evidence tracing and execution provenance in LLM agents, including how retrieved evidence, tool outputs, and memory items influence decisions and outcomes. It proposes a taxonomy, reviews methods and benchmarks, and emphasizes process-level accountability, debugging, and auditability.

• DMF: A Deterministic Memory Framework for Conversational AI Agents 2026-06-02 conf:99% DMF proposes a deterministic memory framework for conversational AI agents that avoids LLM-based summarization when writing and pruning memory. It uses structured signals and decay-based scoring to manage recall efficiently, matching Mem0 accuracy while greatly reducing token usage.

• DELTAMEM: Incremental Experience Memory for LLM Agents via Residual Trees 2026-06-02 conf:99% DeltaMem organizes LLM agent experience into residual trees that store shared base experiences and incremental variations, reducing redundancy and retrieval conflicts. It adds retrieval and consolidation mechanisms so agents can reuse and refine memory more effectively across interactive tasks.

• MemTrain: Self-Supervised Context Memory Training 2026-06-02 conf:98% MemTrain is a self-supervised framework that improves LLM agents’ context memory using masked reconstruction and intermediate memory recall objectives over unlabeled text. It boosts memory-intensive downstream reasoning on long-text and search-based QA benchmarks.

• PhotoCraft: Agentic Reasoning with Hierarchical Self-Evolving Memory for Deep Image Search 2026-06-02 conf:98% PhotoCraft introduces a training-free hierarchical memory system for multimodal LLM agents, combining working, episodic, and semantic memory for deep image search. It improves multi-step reasoning and context-aware retrieval on DISBench, reducing failures caused by stateless agents.

• Better with Experience: Self-Evolving LLM Agents for Evidence-Grounded Health Community Notes 2026-06-01 conf:97% EvoNote is an LLM agent system that improves health Community Notes by storing reusable experience memory from past misinformation correction episodes. It uses fine-grained credit assignment to turn feedback into action-level memory, leading to faster note generation and better evidence-grounded notes than human-written baselines.

• Joint Agent Memory and Exploration Learning via Novelty Signals 2026-06-01 conf:96% JAMEL jointly trains a language-agent memory module and exploration policy using novelty signals as supervision. It compresses long interaction histories into useful latent memory, improving exploration in unseen environments while lowering token usage.

• Latent Collaboration in Multi-Agent Systems 2026-06-01 conf:95% LatentMAS lets LLM agents collaborate through a shared latent working memory instead of text, preserving internal representations without re-encoding. It improves multi-agent reasoning accuracy while greatly reducing token use and speeding up inference.

• Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses 2026-06-01 conf:92% Harness-1 trains a search agent with reinforcement learning while offloading routine state tracking to an external harness that keeps structured working memory. This explicit memory design improves retrieval performance and transfer across benchmarks by separating semantic search decisions from bookkeeping.

• AMP: A Vendor-Neutral Wire Format for Agent Memory Operations 2026-05-31 conf:98% Proposes a vendor-neutral wire format and reference implementation for agent memory operations like remember, recall, forget, and merge across multiple memory frameworks. It adds interoperable memory types, human-in-the-loop governance, and cross-adapter evaluation for practical long-term memory systems.

• Can LLM Agents Sustain Long-Horizon Organizational Dynamics? 2026-05-31 conf:96% TaskWeave is a hierarchical multi-agent framework that uses planning-state maintenance and dependency-aware trace memory to support coherent long-horizon organizational simulations. It improves execution grounding and organizational coherence in a year-long IT company simulation.

• Reasmory: 3D Reconstruction as Explicit Memory for VLMs Spatial Reasoning 2026-05-31 conf:93% Reasmory builds explicit 3D spatial memory from multi-view images or video and lets VLMs reason over it through a validated DSL instead of free-form tool use. This improves spatial reasoning reliability and boosts benchmark performance by 6–18% over strong baselines.

• Don't Read Everything: A Curvature-Conditioned Query for Linear Attention 2026-05-31 conf:92% The paper proposes Curvature-Conditioned Query, a read-time memory access method for linear attention that contracts queries using running key covariance. This improves retrieval and long-context performance with small overhead and can be added to existing linear-attention models.

• MemPro: Agentic Memory Systems as Evolvable Programs 2026-05-30 conf:99% MemPro treats an agent’s memory construction-retrieval pipeline as an evolvable program, not a fixed system. It iteratively diagnoses failures and edits memory-system implementations, improving long-horizon agent performance on multiple memory-intensive benchmarks.

• CoMIC: Collaborative Memory and Insights Circulation for Long-Horizon LLM Agents in Cloud-Edge Systems 2026-05-30 conf:98% CoMIC is a cloud-edge framework that helps lightweight LLM agents handle long-horizon tasks using hierarchical local memory and cloud-based reflection. It shares reusable experience across agents through semantic subgoals, improving task progress and success without fine-tuning model parameters.

• Learning to Retrieve: Dual-Level Long-Term Memory for Text-to-SQL Agents 2026-05-30 conf:98% MERIT is a dual-level long-term memory framework for text-to-SQL agents that retrieves episode-level and turn-level experiences at different interaction stages. It learns retrieval policies with reinforcement learning and dense proxy rewards, improving task success and reducing interaction turns.

• RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates 2026-05-30 conf:95% RedDebate combines multi-agent debate with long-term memory modules that store safety insights from prior interactions and reuse them during later inference. This automated red-teaming framework improves LLM safety by reducing unsafe outputs beyond debate alone.

• SAGE: A Novelty Gate for Efficient Memory Evolution in Agentic LLMs 2026-05-29 conf:99% SAGE is a novelty-aware memory write gate for agentic LLMs that decides when to add, merge, or ignore new facts. It improves long-term memory quality while reducing expensive LLM write-time calls, latency, and cost.

• ElasticMem: Latent Memory as a Learnable Resource for LLM Agents 2026-05-29 conf:99% ElasticMem gives LLM agents a learnable long-term latent memory that adaptively retrieves past information and assigns each memory a variable budget before injecting it as soft tokens. It improves memory-heavy QA and embodied agent performance while reducing token costs.

• AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle 2026-05-29 conf:98% AutoSci is a memory-centric scientific agent system that supports the full research lifecycle using persistent structured memory for reusable knowledge and project artifacts. It combines lifecycle orchestration, multi-agent operators, and feedback-driven memory/skill updates to improve across research projects.

• Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents 2026-05-28 conf:99% The paper proposes MMPO, a method for training memory-augmented LLM agents using a belief-entropy signal that measures how well intermediate summaries preserve task-relevant state. This gives finer-grained supervision than outcome-only rewards and improves long-horizon performance at very large context lengths.

• Hijacking Agent Memory: Stealthy Trojan Attacks Through Conversational Interaction 2026-05-28 conf:98% This paper presents MemPoison, a stealthy attack that injects triggerable backdoors into an LLM agent’s long-term memory through normal conversation. It targets realistic selective memory pipelines and shows high attack success across agents while analyzing why current defenses fail.

• STAMP: Training Explicit Memory for Mobile GUI Agents in Controllable and Scalable Virtual Environments 2026-05-28 conf:98% STAMP trains mobile GUI agents to explicitly encode and retrieve memory using controllable virtual environments that generate scalable, verifiable supervision. It also introduces the Memory-World benchmark and shows strong gains on long-horizon tasks requiring memory.

• MemCollab: Cross-Model Memory Collaboration via Contrastive Trajectory Distillation 2026-05-28 conf:98% MemCollab builds a shared memory for LLM agents with different backbone models by contrasting their reasoning trajectories to extract task-level constraints while filtering model-specific biases. A task-aware retrieval method then selects relevant memories, improving accuracy and efficiency on math and code tasks.

• Entity-Collision: A Stratified Protocol for Attributing Retrieval Lift in Agent Memory 2026-05-28 conf:96% The paper introduces Entity-Collision, a reproducible evaluation protocol for agent-memory retrieval that separates lexical overlap effects from true embedding-based gains. It shows which retrievers help under different memory-query conditions and highlights limits of larger encoders for agent memory recall.

• Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection 2026-05-28 conf:95% Loong is a long-document translation agent that uses a 3E memory module to store summaries, exemplars, and entities, then adaptively selects the most useful history instead of attending to all prior context. It improves translation quality and robustness on ultra-long documents through reinforcement-learned context policies.

• PatchBoard: Schema-Grounded State Mutation for Reliable and Auditable LLM Multi-Agent Collaboration 2026-05-28 conf:94% PatchBoard replaces free-form multi-agent dialogue with validated JSON Patch updates to a shared structured state. This makes LLM agent collaboration more reliable, auditable, and token-efficient, and substantially improves task success on ALFWorld.

• Unlocking the Working Memory of Large Language Models for Latent Reasoning 2026-05-28 conf:93% The paper proposes Reasoning in Memory (RiM), which uses fixed memory blocks to let LLMs perform latent reasoning without generating intermediate thought tokens autoregressively. This enables more compute-efficient reasoning and matches or outperforms existing latent reasoning methods on reasoning benchmarks.

• Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation 2026-05-28 conf:90% Ptah is a multi-agent framework for multimodal deep research that uses a visual working memory to track source-aligned images during planning, evidence collection, and report writing. It also introduces an evaluation protocol for checking factual grounding, citation fidelity, and cross-modal consistency.

• Rethinking Memory as Continuously Evolving Connectivity 2026-05-27 conf:99% FluxMem treats LLM agent memory as a heterogeneous graph that continuously updates its connections through feedback, refinement, and long-term consolidation. It improves adaptation and generalization in dynamic agent tasks, achieving strong results on LoCoMo, Mind2Web, and GAIA.

• MemCog: From Memory-as-Tool to Memory-as-Cognition in Conversational Agents 2026-05-27 conf:99% MemCog turns agent memory from one-shot retrieval into an active part of reasoning through navigable linked memory, multi-step memory traversal, and proactive memory triggering. It also introduces ProactiveMemBench and shows strong gains on both standard and proactive memory benchmarks.

• MemGuard: Preventing Memory Contamination in Long-Term Memory-Augmented Large Language Models 2026-05-27 conf:99% MemGuard is a type-aware long-term memory framework for LLMs that separates user facts, episodic events, and behavioral rules to avoid mixing incompatible memories. It improves reliability in long-horizon conversations while using fewer retrieved memory tokens.

• Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory 2026-05-27 conf:99% BudgetMem is a runtime memory framework for LLM agents that routes queries across memory modules with different cost-performance tiers. It uses a lightweight learned router to control memory construction cost while preserving task accuracy, achieving better trade-offs on several memory-intensive benchmarks.

• Personal Visual Memory from Explicit and Implicit Evidence 2026-05-27 conf:98% Introduces a benchmark for personal visual memory and VisualMem, a hybrid visual-text memory system for personalized AI agents. It captures explicit and implicit user information from images instead of reducing them to captions, improving long-term memory performance.

• ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory 2026-05-27 conf:98% ClinicalAgents introduces a multi-agent clinical decision framework with dual memory: mutable working memory for evolving patient context and static experience memory for retrieving guidelines and past cases. Combined with MCTS-based orchestration, it improves diagnostic accuracy and explainability.

• When Does Memory Help Multi-Trajectory Inference for Tool-Use LLM Agents? 2026-05-27 conf:97% This paper analyzes when memory helps tool-use LLM agents by comparing several cross-trajectory memory methods across multiple inference strategies and benchmarks. It shows that memory gains depend strongly on the search method, with some memory forms improving accuracy only in specific settings and atomic facts mainly reducing trajectory length.

• PersonaAgent: Bridging Memory and Action for Personalized LLM Agents 2026-05-27 conf:97% PersonaAgent is a personalized LLM agent framework that combines episodic and semantic memory with a tool-action module, using persona prompts to connect memory and action. It also introduces a test-time alignment method that updates personalization from recent interactions.

• Hallucination Mitigation with Agentic AI, Nested Learning, and AI Sustainability via Semantic Caching 2026-05-27 conf:94% This paper studies a multi-agent LLM pipeline that uses Continuum Memory Systems and semantic caching to reduce hallucinations without retraining. It shows that layered reviewer agents plus memory-based reuse improve factual reliability, observability, and efficiency.

• Plant, Persist, Trigger: Sleeper Attack on Large Language Model Agents 2026-05-27 conf:89% This paper introduces Sleeper Attack, where adversarial content persists in an LLM agent’s state or memory across interactions and is triggered later by benign queries. It provides a benchmark and experiments showing current LLM agents are vulnerable to these persistent memory-state attacks.

• Agentic Active Omni-Modal Perception for Multi-Hop Audio-Visual Reasoning 2026-05-27 conf:89% The paper presents MOV-Bench for multi-hop audio-visual reasoning and AOP-Agent, an agent framework with hierarchical omni-modal memory and an observe-reflect-replan loop. It improves open-source Omni-LLMs on long-video, reasoning-heavy tasks without extra training.

• Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills 2026-05-27 conf:89% The paper introduces Graph-of-Skills, a structural retrieval layer that selects dependency-aware skill bundles for LLM agents from large skill libraries within limited context budgets. It improves task success while reducing token usage by retrieving executable prerequisite chains instead of loading all skills.

• MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation 2026-05-26 conf:98% The paper presents an LLM agent framework where skills are created, stored, reused, evaluated, and refined over time. It adds skill-level memory that accumulates experience across tasks, improving reuse, adaptation, and overall agent performance.

• Shopping Companion: A Memory-Augmented LLM Agent for Real-World E-Commerce Tasks 2026-05-26 conf:96% Introduces a memory-augmented shopping agent and a benchmark for cross-session preference-aware e-commerce tasks over 1.2M products. It also proposes annotation-free tool-level rewards to improve long-horizon decision-making, leading a lightweight 4B model to outperform stronger baselines.

• VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions 2026-05-26 conf:95% VitaBench 2.0 is a benchmark for testing personalized, proactive LLM agents in long-term user interactions where agents must extract, update, and use user preferences over time. It also provides a memory interface to compare memory architectures and exposes current models’ weaknesses in real-world personalization.

• PEAM: Parametric Embodied Agent Memory through Contrastive Internalization of Experience in Minecraft 2026-05-26 conf:85% PEAM gives a Minecraft agent a two-part memory system: a slow LLM for reasoning and a fast parametric module that internalizes useful experiences as reusable skills. It learns from failure-correction pairs and uses adaptive consolidation rules to improve long-horizon performance while reducing forgetting.

• Memory Architectures for Multi-Turn Text-to-SQL: A Benchmark and Empirical Study 2026-05-25 conf:99% The paper introduces EnterpriseMem-Bench, a multi-turn Text-to-SQL benchmark for testing LLM memory. It empirically compares working memory, episodic retrieval, and semantic augmentation, showing that memory design strongly affects multi-turn accuracy and releasing code and evaluation tools.

• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications 2026-05-25 conf:99% AMA-Bench evaluates long-horizon memory in real and synthetic LLM agent trajectories, exposing failures of existing similarity-retrieval memory systems. The paper also introduces AMA-Agent, which uses a causality graph and tool-augmented retrieval to improve agent memory performance.

• Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems 2026-05-25 conf:98% The paper introduces AgingBench, a benchmark for measuring how deployed LLM agents degrade over time as their memory systems compress, retrieve, revise, and maintain accumulated state. It diagnoses failures across memory pipeline stages and shows that reliable agents need lifespan-aware evaluation and targeted repairs.

• From Model Scaling to System Scaling: Scaling the Harness in Agentic AI 2026-05-25 conf:96% The paper argues that progress in agentic AI depends on scaling the surrounding system harness, not just the model. It highlights trustworthy memory, context governance, dynamic skill routing, and introduces a reference framework plus benchmarks for evaluating long-horizon agent behavior.

• Anticipate and Learn: Unleashing Idle-Time Compute in Proactive Agents 2026-05-25 conf:94% ProAct is a proactive agent that uses dialogue history and persistent memory during idle time to predict users’ likely next needs and gather supporting information in advance. The paper also introduces ProActEval, a benchmark showing better efficiency and fewer hallucinations than reactive agents.

• Dynamic Dual-Granularity Skill Bank for Agentic RL 2026-05-25 conf:94% The paper introduces D2Skill, a dynamic dual-granularity skill bank that stores reusable task-level and step-level experience for agentic RL. It continuously updates, retrieves, and prunes skills from training experience to improve agent performance across multiple benchmarks.

• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents 2026-05-24 conf:99% MemSkill turns fixed LLM agent memory operations into learnable, reusable skills for extracting, consolidating, and pruning interaction history. It uses a controller to select skills and a designer to refine or create new ones, improving memory quality and downstream agent performance.

• MemMorph: Tool Hijacking in LLM Agents via Memory Poisoning 2026-05-24 conf:98% MemMorph shows that attackers can poison an LLM agent’s long-term memory with a few crafted records to steer tool selection. Across multiple benchmarks, agent backbones, and memory modules, it achieves high attack success and highlights memory integrity as a key security weakness.

• APT-Agent: Automated Penetration Testing using Large Language Models 2026-05-24 conf:91% APT-Agent is an automated LLM-based penetration testing system that adds a command-specific memory architecture to retain context across multi-step attacks and a rectification module to fix hallucinated commands. It significantly improves end-to-end exploitation success over prior LLM pentesting agents.

• SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent 2026-05-23 conf:99% SAM is a standalone memory framework for long-horizon reasoning agents that stores compact memory cues plus raw trajectory pages for state-dependent recall. It improves agent performance across several benchmarks without retraining the underlying language model.

• MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection 2026-05-22 conf:98% MemAudit is a post-hoc auditing framework for memory-augmented LLM agents that identifies harmful stored memories using causal influence scoring and structural anomaly detection. It mitigates memory poisoning attacks by tracing which past records caused unsafe behavior and removing their impact.

• OnePred: Next-Query Prediction via Recursive Intent Memory in Multi-Turn Conversations 2026-05-22 conf:97% OnePred predicts users’ next queries by maintaining a recursively updated intent memory instead of rereading full dialogue history. It cuts per-turn token cost substantially, improves prediction quality on long conversations, and introduces NQP-Bench for evaluation.

• Parallel Context Compaction for Long-Horizon LLM Agent Serving 2026-05-22 conf:97% The paper proposes parallel context compaction for long-horizon LLM agents, replacing slow sequential summarization with a faster, more controllable approach. It improves summary-volume predictability and throughput while reducing end-to-end latency on long-context QA and dialogue benchmarks.

• SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills 2026-05-22 conf:93% SkillEvolBench is a benchmark for testing whether LLM agents can turn episodic task trajectories into reusable procedural skills via an external skill library. It shows current agents often fail to form robust transferable skills, with raw trajectory reuse frequently outperforming distilled skills.

• Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning 2026-05-22 conf:88% Agent-ToM monitors autonomous LLM agents by reasoning about their beliefs and intentions over full trajectories. It learns a persistent semantic guardrail memory from prior critiques, reusing constraints across episodes to better detect covert malicious behavior.

• Preisach Attention: A Hysteretic Model of Sequential Memory 2026-05-22 conf:83% The paper introduces Preisach Attention, a hysteresis-based alternative to softmax attention that stores sequence history as a stack of local extrema. It proves strong memory-related properties, including efficient long episodic memory and lower depth requirements for certain history-dependent tasks.

• DeferMem: Query-Time Evidence Distillation via Reinforcement Learning for Long-Term Memory QA 2026-05-21 conf:99% DeferMem is a long-term memory framework for LLM agents that retrieves broad history candidates and then distills query-specific evidence at inference time. It trains the distillation step with reinforcement learning and improves both QA accuracy and memory efficiency on long-memory benchmarks.

• What Training Data Teaches RL Memory Agents: An Empirical Study of Curriculum Effects in Memory-Augmented QA 2026-05-21 conf:98% This paper studies how training curricula affect RL-trained LLM agents that answer questions using external memory in multi-session dialogue. It shows that mixed-benchmark training improves overall performance, while narrow curricula can teach targeted memory skills like temporal reasoning.

• Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention 2026-05-21 conf:93% Gated DeltaNet-2 improves linear-attention memory by separating erase and write operations with channel-wise gates, making compressed recurrent memory easier to update without corrupting stored information. It shows stronger long-context and retrieval performance than related sequence models, especially on needle-in-a-haystack benchmarks.

• Psy-Chronicle:A Structured Pipeline for Synthesizing Long-Horizon Campus Psychological Counseling Dialogues 2026-05-21 conf:88% Psy-Chronicle is a framework for generating long-horizon psychological counseling dialogues using structured memory across sessions. It releases a large Chinese dataset and benchmark to test session response quality, memory recall, and temporal-causal reasoning in counseling agents.

• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective 2026-05-18 conf:99% EvoMemBench is a unified benchmark for evaluating memory in LLM agents across in-episode vs. cross-episode and knowledge- vs. execution-oriented settings. It compares 15 memory methods and long-context baselines, showing current memory systems remain limited and work best when task difficulty or context constraints make memory necessary.

• Remembering More, Risking More: Longitudinal Safety Risks in Memory-Equipped LLM Agents 2026-05-18 conf:99% The paper studies how accumulated memory in LLM agents creates growing safety risks across many independent tasks. It introduces a longitudinal evaluation protocol for detecting memory-induced failures and shows that risk rises with longer memory exposure across multiple memory architectures and agent platforms.

• SocialMemBench: Are AI Memory Systems Ready for Social Group Settings? 2026-05-18 conf:99% SocialMemBench is a benchmark for testing AI memory systems in multi-person social groups, where models must track shared history, norms, and individual exceptions. It reveals major failures in current open-source memory frameworks and shows that social-group memory remains a hard unsolved problem.

• DimMem: Dimensional Structuring for Efficient Long-Term Agent Memory 2026-05-18 conf:99% DimMem introduces a structured long-term memory format for LLM agents that stores atomic memories with explicit fields like time, location, and purpose. This enables more accurate and cheaper retrieval and updating than prior lightweight memory systems on long-context agent benchmarks.

• Hidden in Memory: Sleeper Memory Poisoning in LLM Agents 2026-05-18 conf:99% This paper studies a new attack on LLM agents with persistent memory, where adversarial content poisons stored user memories and later steers future behavior. It evaluates whether poisoned memories are written, retrieved, and acted on, showing persistent memory can become a long-term attack surface.

• Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory 2026-05-18 conf:99% Evo-Memory introduces a benchmark and framework for evaluating self-evolving memory in LLM agents over continuous task streams. It compares memory modules and proposes methods for retrieving, updating, and refining memory so agents can reuse experience and improve at test time.

• EXG: Self-Evolving Agents with Experience Graphs 2026-05-18 conf:97% EXG introduces a structured experience graph as an external memory for self-evolving LLM agents, organizing successes and failures for real-time and offline reuse. It improves cross-task learning, solution quality, and efficiency on code generation and reasoning benchmarks.

• Causal Intervention-Based Memory Selection for Long-Horizon LLM Agents 2026-05-17 conf:99% The paper proposes Causal Memory Intervention, a method that selects long-term memories for LLM agents based on their causal effect on answer quality rather than semantic relevance alone. It also introduces Causal-LoCoMo, a benchmark for testing robustness against irrelevant and harmful memories.

• NeuSymMS: A Hybrid Neuro-Symbolic Memory System for Persistent, Self-Curating LLM Agents 2026-05-17 conf:99% NeuSymMS proposes a hybrid neuro-symbolic memory system for LLM agents that extracts facts from dialogue, curates them with explicit rules, and stores them as persistent structured memory. It supports short- and long-term memory, pruning and promotion policies, and scoped sharing across users and agents.

• MemRepair: Hierarchical Memory for Agentic Repository-Level Vulnerability Repair 2026-05-17 conf:98% MemRepair adds hierarchical persistent memory to LLM repair agents, combining prior fixes, security patterns, and refinement trajectories to improve multi-file vulnerability repair. It uses feedback-driven retrieval and revision, achieving state-of-the-art results on several repository-level repair benchmarks.

• Embodied Task Planning via Graph-Informed Action Generation with Large Language Models 2026-05-17 conf:97% GiG equips embodied LLM agents with a graph-structured experience memory that stores and retrieves similar past execution traces to support long-horizon planning. Combined with bounded symbolic lookahead, it improves grounded action generation and significantly boosts performance on Robotouille and ALFWorld.

• CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery 2026-05-17 conf:92% CORAL is a framework for autonomous LLM multi-agent evolution that uses shared persistent memory, asynchronous execution, and agent management to support long-running open-ended discovery. It shows strong gains on optimization tasks by enabling knowledge reuse and collaboration among agents.

• A Machine With Human-Like Memory Systems 2026-05-17 conf:79% The paper designs an agent with separate semantic and episodic memory systems and shows this outperforms using only one memory type. It also introduces the Room benchmark, where agents learn to store and retrieve memories, including in human-agent and multi-agent collaboration.

• GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations 2026-05-16 conf:99% Introduces GroupMemBench, a benchmark for evaluating LLM agent memory in multi-party conversations. It tests group dynamics, per-user belief tracking, and audience-aware language, showing current memory systems perform poorly and often trail simple BM25 retrieval.

• State Contamination in Memory-Augmented LLM Agents 2026-05-16 conf:98% This paper studies how harmful content can be compressed into seemingly safe memory summaries in LLM agents and still influence later behavior. It introduces the sub-threshold propagation gap (SPG) metric and shows that sanitizing state before summarization is more effective than cleaning summaries afterward.

• When Robots Do the Chores: A Benchmark and Agent for Long-Horizon Household Task Execution 2026-05-16 conf:95% The paper introduces LongAct, a benchmark for long-horizon household tasks, and HoloMind, a VLM agent that uses multimodal spatial memory, episodic memory, and reflective planning. It shows that explicit memory systems improve long-horizon task execution beyond relying on larger models alone.

• FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast 2026-05-15 conf:99% FORGE lets LLM agents improve by turning failed trajectories into reusable text memories like rules or examples, then broadcasting the best memory across agents. On a network-defense benchmark, this memory-sharing scheme beats zero-shot and single-agent reflection baselines without any model weight updates.

• RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents 2026-05-15 conf:99% RecMem is a memory system for long-running LLM agents that delays consolidation until semantically similar interactions recur, cutting token cost. It uses a subconscious retrieval layer plus episodic/semantic memory extraction and refinement, achieving lower cost and higher accuracy than prior memory systems.

• Agentic Recommender System with Hierarchical Belief-State Memory 2026-05-15 conf:99% MARS is a memory-augmented LLM recommender agent that uses hierarchical belief-state memory across events, preferences, and profiles. It introduces adaptive memory lifecycle operations like reinforcement, consolidation, forgetting, and resynthesis, improving personalized recommendation on four benchmarks.

• SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory 2026-05-15 conf:98% SMMBench is a benchmark for testing multimodal agent memory when evidence is scattered across independent sources like chats, images, tables, and documents. It evaluates retrieval, alignment, conflict resolution, preference reasoning, and memory-grounded action prediction, showing current systems still struggle.

• ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts 2026-05-15 conf:98% This paper introduces SHADOWMERGE, a poisoning attack that injects conflicting relations into graph-based agent memory so they are later retrieved and alter LLM agent behavior. It shows high attack success on Mem0 and real-world datasets, exposing security weaknesses in memory-augmented agent systems.

• MemQ: Integrating Q-Learning into Self-Evolving Memory Agents over Provenance DAGs 2026-05-14 conf:99% MemQ improves LLM agent memory by assigning value to retrieved memories through a provenance DAG, letting credit flow backward across memory dependencies. Using Q-learning with eligibility traces, it boosts performance on six benchmarks, especially for multi-step tasks requiring deeper memory chains.

• E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory 2026-05-14 conf:99% E-mem introduces a multi-agent episodic memory system for LLM agents that preserves uncompressed context and reconstructs relevant episodes for reasoning. It uses assistant agents and a master planner to improve long-horizon reasoning, outperforming prior methods on LoCoMo while cutting token costs.

• MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory 2026-05-14 conf:98% MemEye is a visual-centric benchmark for testing long-term multimodal agent memory, especially whether agents retain fine-grained visual evidence and track changing visual states over time. Evaluations across tasks, memory methods, and VLMs show current systems struggle with detail preservation and temporal reasoning.

• MemLineage: Lineage-Guided Enforcement for LLM Agent Memory 2026-05-14 conf:98% MemLineage is a defense for persistent LLM agent memory that tracks cryptographic provenance and derivation lineage for each memory entry. It blocks sensitive actions justified by untrusted memory while preserving benign recall, and shows strong protection against memory-poisoning attacks with minimal overhead.

• MemReranker: Reasoning-Aware Reranking for Agent Memory Retrieval 2026-05-14 conf:98% MemReranker improves long-term memory retrieval in LLM agents by replacing similarity-only reranking with reasoning-aware reranking that handles temporal, causal, and dialogue-context queries. It achieves stronger retrieval accuracy than common rerankers while keeping latency much lower than larger models.

• Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation 2026-05-14 conf:95% MemDocAgent generates repository-level code documentation using a shared external memory that stores and verifies prior agent traces across a dependency-aware traversal. This memory-guided design improves consistency, hierarchy, and overall documentation quality over existing baselines.

• DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery 2026-05-14 conf:93% DrugSAGE gives an LLM-based agent a cross-task memory that stores successful skills, useful strategy patterns, and past mistakes with fixes. This lets the agent reuse experience across drug discovery tasks, reducing search and improving performance, including strong zero-test-time results.

• LiSA: Lifelong Safety Adaptation via Conservative Policy Induction 2026-05-14 conf:89% LiSA improves fixed AI guardrails by turning sparse failure reports into structured reusable memory, with conflict-aware local rules and confidence gating to avoid overgeneralization. It boosts safety adaptation for agents under noisy, limited feedback without repeated fine-tuning.

• Improving Multi-turn Dialogue Consistency with Self-Recall Thinking 2026-05-14 conf:88% The paper proposes Self-Recall Thinking, a framework that teaches LLMs to selectively recall relevant past dialogue turns and reason over them during inference. It improves multi-turn dialogue consistency while reducing latency compared with external-memory and summarization-based methods.

• Cognifold: Always-On Proactive Memory via Cognitive Folding 2026-05-13 conf:99% Cognifold introduces an always-on, brain-inspired memory system for agents that proactively organizes event streams into evolving cognitive structures. It adds an intent layer to complementary learning systems and shows strong results on both structural memory formation and standard memory benchmarks.

• Useful Memories Become Faulty When Continuously Updated by LLMs 2026-05-13 conf:99% The paper shows that continuously consolidated textual memories in LLM agents often become faulty and can hurt performance, even when built from correct experiences. It finds episodic memory retention is more reliable than frequent consolidation and argues consolidation should be explicitly gated.

• BOOKMARKS: Efficient Active Storyline Memory for Role-playing 2026-05-13 conf:98% BOOKMARKS introduces a search-based memory framework for role-playing agents that stores and updates task-relevant storyline bookmarks instead of relying on repeated summarization. It improves long-horizon consistency by actively grounding details and passively synchronizing memory more efficiently.

• PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records 2026-05-13 conf:98% PersonalAlign studies personalized GUI agents that use long-term user records as persistent memory to infer implicit preferences and routines. It introduces the AndroidIntent benchmark and HIM-Agent, a hierarchical, continuously updated memory system that improves execution and proactive assistance.

• ScioMind: Cognitively Grounded Multi-Agent Social Simulation with Anchoring-Based Belief Dynamics and Dynamic Profiles 2026-05-13 conf:86% ScioMind is an LLM multi-agent social simulation framework that adds hierarchical memory, memory-anchored belief updates, and retrieval-based dynamic profiles to produce more stable and realistic belief trajectories. It shows that memory and reflection improve behavioral realism in policy debate simulations.

• $δ$-mem: Efficient Online Memory for Large Language Models 2026-05-12 conf:99% The paper introduces δ-mem, a lightweight online memory that adds a compact associative state to a frozen LLM and uses it to adjust attention during generation. It improves performance on memory-heavy benchmarks without extending context windows or fully fine-tuning the backbone.

• MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare 2026-05-12 conf:98% MedMemoryBench is a benchmark for testing memory in personalized healthcare agents using realistic long-horizon medical interactions. It introduces a streaming evaluation protocol and studies memory saturation, showing current systems struggle with retrieval, reasoning, and noise over time.

• Automated Reformulation of Robust Optimization via Memory-Augmented Large Language Models 2026-05-12 conf:91% The paper introduces AutoREM, a tuning-free memory-augmented framework that builds structured experience memory from past failed reformulation attempts to improve robust optimization reformulation by LLMs. It also presents AutoRO-Bench, a benchmark for systematically evaluating this task.

• Nautilus Compass: Black-box Persona Drift Detection for Production LLM Agents 2026-05-11 conf:99% Nautilus Compass is a black-box memory layer for production LLM agents that detects persona drift and retrieves relevant past conversation text without requiring model access or LLM-based fact extraction. It provides a low-cost, deployable memory system with benchmarked gains on long-context and agent memory tasks.

• PREPING: Building Agent Memory without Tasks 2026-05-11 conf:98% Preping builds agent procedural memory before real tasks by generating synthetic practice and selectively storing useful trajectories. It uses a proposer, solver, and validator to control what gets practiced and remembered, improving performance while lowering deployment cost.

• Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory 2026-05-11 conf:98% The paper frames agent memory as a rate-distortion problem, where good memory preserves distinctions needed for decisions rather than detailed descriptions. It proposes DeMem, an online memory learner that forgets safely under a fixed budget and improves long-horizon agent performance.

• The Trap of Trajectory: Towards Understanding and Mitigating Spurious Correlations in Agentic Memory 2026-05-10 conf:99% The paper studies how agentic memory can amplify spurious correlations in LLM decision-making. It introduces a benchmark for trajectory-level memory failures and proposes CAMEL, a plug-and-play calibration method that reduces harmful memory reliance during writing and retrieval.

• EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium 2026-05-10 conf:99% EquiMem improves shared memory in multi-agent debate by treating memory updates as a zero-trust game and calibrating them at inference time without relying on extra LLM judgments. It strengthens robustness against corrupted or adversarial memory entries across different memory architectures.

• Portable Agent Memory: A Protocol for Cryptographically-Verified Memory Transfer Across Heterogeneous AI Agents 2026-05-10 conf:98% This paper proposes an open protocol for transferring persistent memory across different AI agents and models. It introduces a structured memory format with provenance tracking, selective access control, and safe rehydration, plus an open-source SDK demonstrating cross-model memory transfer.

• Mem-W: Latent Memory-Native GUI Agents 2026-05-10 conf:98% Mem-W introduces latent memory-native GUI agents that store trajectory and in-session experience as compact memory tokens integrated directly into the model’s embedding context. This avoids text-based memory scaffolds and improves long-horizon GUI task performance across web and mobile benchmarks.

• MEMOREPAIR: Barrier-First Cascade Repair in Agentic Memory 2026-05-08 conf:99% The paper introduces MemoRepair, a method for safely repairing agent memory when source artifacts change or become invalid. It withdraws stale descendants, rebuilds valid successors under current interfaces, and uses an exact min-cut formulation to optimize repair cost while eliminating invalidated-memory exposure.

• MemSearcher: Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learning 2026-05-08 conf:99% MemSearcher trains LLM search agents to keep a compact, question-relevant memory instead of concatenating full interaction history. It introduces multi-context GRPO for end-to-end reinforcement learning across multi-turn trajectories, improving search performance while keeping context length nearly constant.

• A Multi-Memory Segment System for Generating High-Quality Long-Term Memory Content in Agents 2026-05-08 conf:99% The paper proposes a cognitively inspired multi-memory segment system that converts short-term interactions into structured long-term memory units for better retrieval and response generation in agents. Experiments on LoCoMo show improved memory quality, robustness, and practical effectiveness.

• PYTHALAB-MERA: Validation-Grounded Memory, Retrieval, and Acceptance Control for Frozen-LLM Coding Agents 2026-05-08 conf:98% PYTHALAB-MERA adds an external memory-and-retrieval controller to frozen local coding LLMs, selecting episodic records and reusable skills for prompting based on validation feedback. It uses bounded rewards and delayed credit assignment to improve strict task validation success in coding-agent settings.

• When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory 2026-05-08 conf:98% This paper proposes a scale-conditioned evaluation protocol for agent memory that tests whether stored evidence remains usable as irrelevant sessions accumulate. It introduces diagnostics for reliability, memory-call burden, failure modes, and usable scale, and applies them to benchmark memory interfaces and agents.

• Why Retrying Fails: Context Contamination in LLM Agent Pipelines 2026-05-08 conf:96% The paper formalizes how failed attempts left in an LLM agent’s context window degrade later retries, raising error rates in multi-step tool-use pipelines. It derives closed-form results for success probability and retry budgets, and shows on SWE-bench Verified that clearing context before retry can substantially improve outcomes.

• The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents 2026-05-08 conf:93% This paper shows that giving LLM agents longer recall can reduce cooperation in repeated multi-agent games, a phenomenon it calls the memory curse. It traces the effect to memory content and reasoning patterns, and shows cooperation can be restored through memory sanitization or fine-tuning on forward-looking reasoning.

• STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? 2026-05-07 conf:99% STALE is a benchmark for testing whether LLM agents can detect and revise outdated memories when new evidence implicitly invalidates prior beliefs. It also introduces CUPMem, a prototype memory system that improves state-aware memory updating through structured consolidation and propagation-aware search.

• What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis 2026-05-07 conf:98% This paper analyzes how LLM agents internally implement memory write and read operations across external memory systems. It finds distinct routing and content circuits, shows memory uses existing context-grounding circuitry in the base model, and introduces an unsupervised diagnostic to localize memory failures.

• LMEB: Long-horizon Memory Embedding Benchmark 2026-05-07 conf:98% LMEB is a benchmark for testing embedding models on long-horizon memory retrieval in memory-augmented LLM systems. It covers 22 datasets and 193 zero-shot tasks across episodic, dialogue, semantic, and procedural memory, showing that standard retrieval benchmarks miss important memory-related capabilities.

• Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks 2026-05-07 conf:98% The paper introduces MemAct, a framework that treats LLM working-memory management as learnable actions like deleting and inserting context during long-horizon agent tasks. It jointly optimizes memory retention and task performance with reinforcement learning, reducing context length while maintaining strong accuracy.

• SkillOS: Learning Skill Curation for Self-Evolving Agents 2026-05-07 conf:96% SkillOS trains a skill curator with reinforcement learning to update an external skill repository that LLM agents retrieve from over task streams. It improves long-term skill curation and agent performance, outperforming memory-free and prior memory-based baselines across tasks.

• More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding 2026-05-07 conf:91% This paper shows that adding more agent components, including memory, does not always improve LLM agent performance because components can interfere with each other. It finds task- and model-dependent best subsets and argues for interaction-aware selection instead of default all-in agent stacks.

• CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment 2026-05-05 conf:99% CASCADE gives LLM agents an explicit episodic memory that stores and reuses past cases during deployment. It frames memory retrieval and updating as a contextual bandit, enabling continual adaptation without finetuning and improving performance across a wide range of tasks.

• MAGE: Safeguarding LLM Agents against Long-Horizon Threats via Shadow Memory 2026-05-04 conf:98% MAGE adds a dedicated shadow memory to LLM agents that stores safety-critical context across long task trajectories and checks pending actions for risk. It improves detection and early mitigation of long-horizon attacks with little overhead.

• The Dynamic Gist-Based Memory Model (DGMM): A Memory-Centric Architecture for Artificial Intelligence 2026-05-04 conf:98% The paper proposes DGMM, a memory-centric AI architecture that stores experience in an explicit, evolving graph of episodic and semantic memory. It uses cue-based recall to build working memory, aiming to improve persistence, provenance, interpretability, and temporal grounding without retraining.

• A Semantic Autonomy Framework for VLM-Integrated Indoor Mobile Robots: Hybrid Deterministic Reasoning and Cross-Robot Adaptive Memory 2026-05-04 conf:96% This paper proposes a semantic autonomy framework for VLM-powered indoor robots that combines fast deterministic reasoning with an explicit adaptive memory shared across sessions and robots. Its memory design stores preferences, environment knowledge, and robot capabilities, enabling cross-robot transfer and large latency reductions on edge hardware.

• WMF-AM: Probing LLM Working Memory via Depth-Parameterized Cumulative State Tracking 2026-05-03 conf:89% WMF-AM is a benchmark for measuring LLM working memory by testing how well models maintain and update intermediate state across sequential operations in one pass. It provides a recalibratable diagnostic showing where models break down under cumulative memory load.

• Feedback-Normalized Developer Memory for Reinforcement-Learning Coding Agents: A Safety-Gated MCP Architecture 2026-05-02 conf:98% The paper proposes a safety-gated persistent memory system for RL coding agents that logs retrieval decisions, incorporates feedback, and links verified fixes to prior memory use. It evaluates this auditable MCP-native architecture on a benchmark of RL software bugs and shows strong retrieval reliability with conservative deployment controls.

• Agentic Learner with Grow-and-Refine Multimodal Semantic Memory 2026-05-02 conf:98% This paper introduces ViLoMem, a dual-stream multimodal semantic memory for agentic MLLMs that stores and updates visual and reasoning knowledge separately. It helps models learn from past successes and failures, reducing repeated errors and improving accuracy across multimodal benchmarks.

• $How^{2}$: How to learn from procedural How-to questions 2026-05-02 conf:97% Introduces How^2, a memory framework where LLM-based agents ask how-to questions, store the answers, and reuse them for lifelong learning in interactive environments. In Minecraft planning tasks, abstract reusable answers improve long-term planning more than state-specific executable advice.

• From Unstructured Recall to Schema-Grounded Memory: Reliable AI Memory via Iterative, Schema-Aware Extraction 2026-05-01 conf:99% The paper proposes a schema-grounded external memory system for AI agents that writes verified structured records instead of relying on text retrieval alone. Its iterative, schema-aware extraction pipeline improves reliability on memory tasks requiring exact facts, state updates, and constrained querying.

• Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework 2026-05-01 conf:99% The paper unifies existing memory methods for LLM-based agents, compares them under shared experimental settings, and analyzes their strengths on standard benchmarks. It also proposes a new modular memory method that outperforms prior approaches.

• Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory 2026-05-01 conf:98% MemCoE is a two-stage, cognition-inspired framework for LLM agents that learns both how to organize memory and what user information to update over long interactions. It improves personalized long-term memory with more stable and robust optimization across multiple benchmarks.

• Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents 2026-04-30 conf:99% AgeMem lets an LLM agent manage short-term and long-term memory itself by choosing when to store, retrieve, update, summarize, or discard information. It trains this unified memory behavior with reinforcement learning and improves long-horizon agent performance and context efficiency.

• Contextual Agentic Memory is a Memo, Not True Memory 2026-04-30 conf:98% The paper argues that current agent memory systems are retrieval tools, not true memory, and formalizes their limits in long-term learning, compositional generalization, and security. It proposes a neuroscience-inspired view where retrieval should be paired with slower weight consolidation.

• Grounding Agent Memory in Contextual Intent 2026-04-30 conf:98% STITCH is an agent memory system that tags interaction history with contextual intent so LLM agents can retrieve past information that matches the current goal and action. It reduces retrieval interference in long-horizon tasks and is evaluated with a new benchmark, CAME-Bench.

• OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory 2026-04-29 conf:99% OCR-Memory stores long agent trajectories as images and retrieves relevant verbatim text through visual anchors, reducing token costs and hallucination. It improves long-horizon agent performance under tight context limits by expanding effective external memory capacity.

• Detecting Clinical Discrepancies in Health Coaching Agents: A Dual-Stream Memory and Reconciliation Architecture 2026-04-29 conf:98% Introduces a dual-stream memory system for healthcare LLM agents that keeps patient-reported memories separate from clinical records and reconciles conflicts for safety. It shows strong discrepancy detection across longitudinal sessions and quantifies errors introduced during memory extraction from conversations.

• When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents 2026-04-29 conf:98% This paper shows that in memory-augmented LLM agents, continual learning bottlenecks shift from parameter updates to memory retrieval under limited context. It introduces a framework for memory representation and organization, finding trade-offs between transfer, forgetting, and negative transfer across sequential tasks.

• From Soliloquy to Agora: Memory-Enhanced LLM Agents with Decentralized Debate for Optimization Modeling 2026-04-28 conf:97% Agora-Opt is a multi-agent framework for optimization modeling that combines decentralized debate with a read-write memory bank. It stores solver-verified artifacts and past resolutions so agents can improve over time without training, leading to stronger performance across benchmarks and LLM backbones.

• PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor 2026-04-28 conf:97% PsychAgent is a lifelong-learning counseling agent that uses persistent memory and strategic planning to maintain continuity across multi-session conversations. It also extracts skills from past counseling trajectories and internalizes them to improve future responses.

• MemRec: Collaborative Memory-Augmented Agentic Recommender System 2026-04-28 conf:97% MemRec introduces a collaborative memory framework for LLM-based recommender agents that shares relational signals across users and items while avoiding context overload. It uses a lightweight model to manage and distill dynamic memory for a larger recommendation model, improving performance on four benchmarks.

• Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents 2026-04-23 conf:99% Memanto is a memory layer for long-horizon AI agents that uses typed semantic memory, conflict resolution, and temporal versioning with fast information-theoretic retrieval. It outperforms graph- and vector-based memory systems on LongMemEval and LoCoMo while reducing retrieval and ingestion complexity.

• StructMem: Structured Memory for Long-Horizon Behavior in LLMs 2026-04-23 conf:99% StructMem is a hierarchical memory system for long-term LLM agents that preserves event relationships for better temporal reasoning and multi-hop question answering. It improves performance on LoCoMo while using fewer tokens, API calls, and runtime than prior memory methods.

• FSFM: A Biologically-Inspired Framework for Selective Forgetting of Agent Memory 2026-04-23 conf:99% This paper proposes a biologically inspired framework for selective forgetting in LLM agent memory systems. It organizes forgetting methods into several types and shows that pruning or deleting memories can improve efficiency, response quality, and security in memory-enabled agents.

• AEL: Agent Evolving Learning for Open-Ended Environments 2026-04-23 conf:97% AEL helps LLM agents improve over long sequential tasks by learning which memory retrieval policy to use and by reflecting on failures to guide future decisions. It shows that better use of experience, not extra architectural complexity, drives stronger agent performance.

• Ask Only When Needed: Proactive Retrieval from Memory and Skills for Experience-Driven Lifelong Agents 2026-04-22 conf:99% This paper introduces ProactAgent, a lifelong agent framework that stores past experience in structured factual, episodic, and skill memories and learns to proactively retrieve from them during interaction. It treats retrieval as a decision policy, training the agent to ask for memory only when it improves outcomes or efficiency, leading to better long-horizon task performance with less retrieval overhead.

• HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents 2026-04-22 conf:99% HiGMem is a hierarchical long-term memory system for conversational LLM agents that stores event summaries and dialogue turns, then uses the LLM to decide which detailed turns to read. This improves retrieval precision and answer quality on LoCoMo10 while using far fewer retrieved memory turns than prior methods.

• Lightweight LLM Agent Memory with Small Language Models 2026-04-22 conf:99% LightMem is a lightweight external memory system for LLM agents that uses small language models to manage short-term, mid-term, and long-term memory. It separates fast online retrieval from offline consolidation, aiming to improve long-horizon consistency and reusable knowledge recall with low latency.

• Stateless Decision Memory for Enterprise AI Agents 2026-04-22 conf:98% The paper proposes Deterministic Projection Memory, a stateless memory design for enterprise decision agents that stores an append-only event log and builds a task-specific projection only at decision time. It argues this architecture better supports replayability, auditability, isolation, and scale than stateful summarization-based memory, while improving accuracy and speed under tight memory budgets.

• Memory-Augmented LLM-based Multi-Agent System for Automated Feature Generation on Tabular Data 2026-04-22 conf:88% This paper presents MALMAS, a multi-agent LLM system for automated feature generation on tabular data. It uses explicit procedural, feedback, and conceptual memory modules plus a router agent to iteratively refine generated features and improve their quality and diversity.

• Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents 2026-04-22 conf:86% This paper studies LLM agents playing repeated Avalon games while remembering prior interactions across games. The retained memory produces reputation effects and more strategic deception, showing how agent memory influences multi-agent social dynamics over time.

• IMPACT-CYCLE: A Contract-Based Multi-Agent System for Claim-Level Supervisory Correction of Long-Video Semantic Memory 2026-04-22 conf:80% IMPACT-CYCLE presents a multi-agent system that represents long-video understanding as a shared, versioned semantic memory made of typed claims, dependencies, and provenance. Specialized agents and human arbitration iteratively verify and correct only affected memory regions, improving downstream reasoning while reducing human correction cost.

• From Verbatim to Gist: Distilling Pyramidal Multimodal Memory via Semantic Information Bottleneck for Long-Horizon Video Agents 2026-04-21 conf:98% The paper proposes MM-Mem, a pyramidal multimodal memory system for long-horizon video agents that organizes information into sensory, episodic, and symbolic memory layers. It compresses detailed observations into higher-level semantic representations and uses learned retrieval to access relevant memory, improving long-context video reasoning.

• CLAG: Adaptive Memory Organization via Agent-Driven Clustering for Small Language Model Agents 2026-04-20 conf:99% CLAG is a memory framework for small language model agents that organizes stored experiences into semantic clusters using an agent-driven router. It builds summaries and tags for each cluster and retrieves memory in two stages to reduce irrelevant context, improving QA performance and robustness.

• Latent Preference Modeling for Cross-Session Personalized Tool Calling 2026-04-20 conf:97% This paper studies how LLM agents can remember user preferences across sessions to improve personalized tool calling. It introduces the MPT benchmark and proposes PRefine, a memory-augmented test-time method that extracts and updates reusable preference hypotheses from past interactions, improving accuracy with far less context than full-history prompting.

• HELM: Harness-Enhanced Long-horizon Memory for Vision-Language-Action Manipulation 2026-04-20 conf:96% HELM improves long-horizon vision-language-action manipulation by adding an episodic memory module that retrieves key task history, a learned verifier that uses memory-conditioned context to predict failures, and a controller that rolls back and replans. The paper shows explicit memory helps more than just extending context length, and introduces an evaluation protocol for recovery under perturbations.

• Memory Intelligence Agent 2026-04-19 conf:98% This paper proposes Memory Intelligence Agent, a deep research agent framework with both non-parametric and parametric memory. It stores compressed historical trajectories, updates memory during test time, and links external and internal memory through a bidirectional conversion loop to improve planning and reasoning.

• Visual Inception: Compromising Long-term Planning in Agentic Recommenders via Multimodal Memory Poisoning 2026-04-18 conf:96% This paper shows that agentic recommender systems using long-term multimodal memory can be attacked by planting poisoned visual memories that later get retrieved and alter planning. It proposes CognitiveGuard, a defense that sanitizes perceptual inputs and verifies memory-based reasoning to reduce the attack’s success.

• Oblivion: Self-Adaptive Agentic Memory Control through Decay-Driven Activation 2026-04-17 conf:99% This paper introduces Oblivion, a memory control framework for LLM agents that uses decay-driven forgetting instead of always-on retrieval. It separates when the agent should read from memory and what memories should be reinforced on write, enabling hierarchical memory and better adaptation in long-horizon tasks.

• Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents 2026-04-17 conf:98% This paper studies how LLM agents can store and compress past experience for long-horizon, multi-session use. It unifies memory, skills, and rules as different compression levels of reusable experience, surveys existing systems, and identifies open problems in adaptive compression and knowledge lifecycle management.

• EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems 2026-04-17 conf:84% The paper introduces J-TTL, a benchmark where an agent must improve across repeated game episodes, and proposes EvoTest, which evolves the agent system after each episode without fine-tuning. EvoTest updates prompts, logs useful state-action choices into memory, adjusts hyperparameters, and refines tool-use routines to improve future performance.

• MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents 2026-04-16 conf:99% MAGMA is an agentic memory architecture that stores memories in multiple relational graphs, including semantic, temporal, causal, and entity views. It retrieves information through policy-guided graph traversal to build more relevant and interpretable context for long-horizon reasoning. Experiments on LoCoMo and LongMemEval show improved performance over prior agent memory systems.

• Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory 2026-04-16 conf:98% This paper introduces Rashomon Memory, a multi-perspective agent memory system where different goal-conditioned agents store the same experiences with different interpretations. At retrieval time, these perspectives debate using argumentation, allowing the system to select, combine, or explicitly surface conflicting memories with an explanation trace.

• PolicyBank: Evolving Policy Understanding for LLM Agents 2026-04-16 conf:97% PolicyBank is a memory system for LLM agents that stores structured policy insights and updates them through corrective feedback, helping agents refine how they interpret organizational policies during tool use. The paper also adds a benchmark setting with controlled policy gaps to evaluate whether memory mechanisms can improve policy compliance beyond treating the original policy text as fixed ground truth.

• Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents 2026-04-16 conf:97% Skill-Pro enables LLM agents to turn past interaction trajectories into reusable procedural skills without changing model weights. It builds a compact long-term procedural memory with mechanisms for skill extraction, verification, reuse, and compression, leading to better performance and stability across tasks and agents.

• APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI 2026-04-15 conf:99% APEX-MEM is a long-term conversational memory system for LLM-based dialogue agents. It stores conversations in a temporally grounded property graph with append-only history and uses a retrieval agent to resolve evolving or conflicting facts at query time, producing compact relevant memory summaries. The system improves performance on LOCOMO and LongMemEval, showing stronger temporal coherence in conversational memory.

• AMA: Adaptive Memory via Multi-Agent Collaboration 2026-04-15 conf:99% AMA is a multi-agent memory framework for LLM agents that builds and retrieves memories at multiple granularities, then checks and refreshes them to maintain relevance and consistency. It explicitly addresses memory construction, adaptive access, conflict resolution, and forgetting/updating, and shows strong results on long-context benchmarks with much lower token use than full-context methods.

• Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution 2026-04-15 conf:99% This paper presents ReMe, a procedural memory framework for LLM agents that turns past experience into reusable memories. It distills successes and failures, retrieves context-relevant memories for new tasks, and updates or prunes memories based on utility. Experiments on BFCL-V3 and AppWorld show strong gains, suggesting dynamic memory can improve agent performance more efficiently than simply scaling model size.

• Memp: Exploring Agent Procedural Memory 2026-04-15 conf:99% The paper proposes Memp, a procedural memory system for LLM agents that distills past trajectories into detailed instructions and higher-level scripts. It explicitly studies how to build, retrieve, update, correct, and deprecate memory over time, showing improved agent performance and transfer of learned procedural memory from stronger to weaker models.

• Evo-MedAgent: Beyond One-Shot Diagnosis with Agents That Remember, Reflect, and Improve 2026-04-15 conf:98% Evo-MedAgent equips a tool-using medical LLM agent with a self-evolving memory so it can learn across cases at test time instead of treating each case independently. Its memory includes retrieved past clinical episodes, an updated bank of diagnostic heuristics, and a controller that tracks tool reliability, improving diagnostic accuracy without retraining.

• Memory Transfer Learning: How Memories are Transferred Across Domains in Coding Agents 2026-04-15 conf:97% This paper studies how coding agents can reuse explicit memories across different task domains instead of keeping memory siloed within one domain. It shows that abstract, high-level memories transfer better than low-level traces, that larger shared memory pools help, and that memory can even transfer across different models.

• ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models 2026-04-15 conf:97% This paper introduces ImplicitMemBench, a benchmark for testing implicit memory in LLM agents through procedural memory, priming, and classical conditioning tasks. Across 17 models, results show that current systems struggle to turn experience into automatic behavior, suggesting important gaps in LLM memory mechanisms and evaluation.

Retrieval-Augmented Generation

RAG pipelines, grounding, and external knowledge retrieval to augment LLM generation.

• Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads 2026-06-04 conf:99% This paper characterizes LLM agent memory systems for long-horizon tasks, introducing a taxonomy and profiling framework to measure construction, retrieval, and generation costs. It evaluates ten systems and distills practical design recommendations for efficient, scalable agent memory.

• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents 2026-06-04 conf:99% The paper introduces MRAgent, an LLM agent framework that stores memories in an associative graph and actively reconstructs relevant paths during reasoning. This dynamic memory access improves long-context reasoning accuracy while reducing token and runtime costs.

• SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents 2026-06-04 conf:99% SubtleMemory is a benchmark for testing whether long-horizon AI agents can preserve and use fine-grained relations among long-term memories, including complementary and conflicting information. It also provides diagnostic protocols to analyze failures in memory storage, retrieval, and downstream reasoning.

• AdaMEM: Test-Time Adaptive Memory for Language Agents 2026-06-04 conf:99% AdaMEM gives language agents a hybrid memory system that combines offline long-term experience storage with dynamically generated short-term strategy memory during execution. This lets agents adapt continuously at test time and improves performance on ALFWorld, WebShop, and HotpotQA.

• Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval 2026-06-04 conf:97% The paper studies an agent-orchestrated adaptive RAG system that uses query decomposition, iterative retrieval, and self-reflection for complex queries. It shows these memory-access enhancements help in some domains but can hurt multi-hop retrieval precision, arguing for selective, cost-aware use.

• FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG 2026-06-04 conf:97% FIDES is a training-free decoding method for RAG that detects token-level conflicts between retrieved evidence and a model’s internal memory. It uses multiple internal signals to apply selective intervention during generation, improving faithfulness and answer quality across several benchmarks and model sizes.

• Self-Augmenting Retrieval for Diffusion Language Models 2026-06-04 conf:96% SARDI is a training-free RAG method for diffusion language models that uses low-confidence intermediate tokens as lookahead cues to retrieve evidence during denoising. It improves multi-hop QA performance over existing diffusion and autoregressive retrieval baselines while running much faster.

• Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation (long version) 2026-06-04 conf:94% The paper proposes a simple graph-based RAG system with agent tools combining vector search and graph queries over structured Wikipedia data. On the MoNaCo benchmark, it improves factual correctness and roughly halves hallucinations with only modest extra token cost.

• RAG Security and Privacy: Formalizing the Threat Model and Attack Surface 2026-06-04 conf:89% This paper introduces a formal threat model for retrieval-augmented generation systems, defining adversaries and attack vectors such as membership inference and data poisoning. It provides a structured foundation for analyzing privacy and security risks in RAG-based LLM memory systems.

• Exploring Cross-Scenario Generality of Agentic Memory Systems: Diagnostics and a Strong Baseline 2026-06-03 conf:99% The paper evaluates eight LLM agent memory systems across diverse scenarios and finds that active, agent-controlled storage and retrieval generalize better than passive memory pipelines. It introduces AutoMEM, a self-managed tool-based memory harness that achieves the strongest cross-scenario performance.

• MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning 2026-06-03 conf:98% Introduces MemoryDocDataSet, a benchmark for testing LLMs on combined conversational memory and long-document reasoning. It shows that standard long-context and document-only retrieval methods struggle on questions requiring both memory of prior conversations and document navigation, motivating unified memory-retrieval systems.

• Cascading Hallucination in Agentic RAG: The CHARM Framework for Detection and Mitigation 2026-06-03 conf:91% The paper defines cascading hallucination in multi-step agentic RAG and introduces CHARM, a framework that detects and interrupts error propagation across reasoning stages. It improves reliability with stage-level verification, cross-stage consistency checks, and confidence monitoring.

• Revisiting Vul-RAG: Reproducibility and Replicability of RAG-based Vulnerability Detection with Open-Weight Models 2026-06-03 conf:86% This paper reproduces and extends evaluation of Vul-RAG, a retrieval-augmented vulnerability detection framework, using fully local open-weight LLMs. It finds the original results largely reproducible but shows performance plateaus across newer and larger models, suggesting model scale alone does not improve RAG-based detection.

• Core-based Hierarchies for Efficient GraphRAG 2026-06-02 conf:96% This paper improves GraphRAG by replacing unstable Leiden clustering with deterministic k-core hierarchies for organizing retrieved knowledge. The method builds efficient, token-budget-aware graph communities that improve answer quality and diversity while lowering LLM token costs.

• When Retrieval Doesn't Help: A Large-Scale Study of Biomedical RAG 2026-06-02 conf:95% This paper evaluates biomedical RAG at scale across multiple models, datasets, retrievers, and corpora, finding that retrieval gives only small, inconsistent gains. It argues the main limitation is not retrieval quality, but LLMs' weak ability to use retrieved evidence effectively.

• Ranking Free RAG: Replacing Re-ranking with Selection in RAG for Sensitive Domains 2026-06-02 conf:95% METEORA improves RAG by replacing opaque re-ranking with rationale-based evidence selection and adaptive cutoff detection. It also uses a verifier to filter poisoned evidence, yielding better retrieval quality, stronger robustness, and less evidence overhead across six datasets.

• Unified Context Evolution for LLM Agents 2026-06-01 conf:99% The paper proposes Unified Context Evolution, a gradient-free memory framework for LLM agents that stores typed experiences as reusable context units. It retrieves, scores, prunes, and balances different memory types over time, substantially improving interactive task performance and enabling transfer without retraining.

• AGENTCL: Toward Rigorous Evaluation of Continual Learning in Language Agents 2026-06-01 conf:98% AGENTCL introduces a benchmark for rigorously evaluating continual learning in language agents using controlled task streams with reusable prior experience. It also presents MemProbe to analyze non-parametric memory designs and shows how memory choices affect transfer, reuse, and degradation across tasks.

• When Knowledge Is Not Free: Cost-Aware Evidence Selection in Retrieval-Augmented Generation 2026-06-01 conf:97% This paper introduces cost-aware RAG, where external evidence has access costs and systems must answer under a budget. It benchmarks budgeted evidence selection and shows that agentic LLMs can adaptively decide when and what to retrieve, though performance is task- and model-dependent.

• Self-Conditioned Positional HNSW for Overlap-Aware Retrieval in Chunked-Document RAG Systems: Method and Industrial Evidence-Quality Audit 2026-06-01 conf:93% Proposes SCP-HNSW, an overlap-aware retrieval method for chunked-document RAG that reduces redundant adjacent chunks by adding positional signals and a query-specific prior. It preserves standard HNSW indexing while improving context selection and auditability for evidence quality.

• TechGraphRAG: An Agentic Graph-Augmented RAG Framework for Technical Literature Reasoning 2026-06-01 conf:92% TechGraphRAG is an agentic graph-augmented RAG system for technical literature reasoning that iteratively searches external sources, traverses a knowledge graph, scores evidence sufficiency, and verifies citations. It shows how a self-correcting, evidence-grounded memory pipeline can improve reasoning over domain-specific corpora.

• Seeing Through the MiRAGE: Evaluating Multimodal Retrieval Augmented Generation 2026-06-01 conf:92% MiRAGE is a framework for evaluating multimodal retrieval-augmented generation using claim-based measures of factual coverage and citation quality. It provides both human-aligned and automatic metrics, plus open-source tools, to assess RAG systems that use audiovisual as well as text sources.

• MemoNoveltyAgent: A Historical Research Memory-Aware Agent Workflow for Paper Novelty Assessment 2026-06-01 conf:91% This paper presents a multi-agent LLM workflow for paper novelty assessment that combines RAG with a hierarchical historical research memory built from scholarly corpora. It also introduces fine-grained novelty-point analysis and a RAG-based checklist evaluation to improve faithfulness and evidence grounding.

• Don't Ask the LLM to Track Freshness: A Deterministic Recipe for Memory Conflict Resolution 2026-05-31 conf:98% This paper shows that LLM memory failures on changing facts often come from post-retrieval conflict resolution, not storage. It proposes a deterministic version-aware aggregation method that selects the newest fact by serial or timestamp, substantially improving results on MemoryAgentBench and related memory-update evaluations.

• ExpWeaver: LLM Agents Learn from Experience via Latent RAG 2026-05-31 conf:98% ExpWeaver lets LLM agents learn from past experiences through latent-space retrieval instead of adding retrieved text to the prompt. It integrates experience memory directly into decoding, improving accuracy, token efficiency, and cross-domain generalization across 13 tasks.

• Efficient RAG with Intent-Aware Retrieval and Semantics-Preserving Chunking 2026-05-31 conf:97% This paper proposes InSemRAG, a retrieval-augmented generation framework that improves LLM access to external knowledge through intent-aware retrieval and semantics-preserving chunking. It uses an iterative retrieve-and-check process with small language models to boost evidence quality while reducing latency.

• Bridging Requirements and Architecture: Multi-Agent Orchestration with External Knowledge and Hierarchical Memory 2026-05-31 conf:95% MAAD is a multi-agent framework for software architecture design that combines RAG with hierarchical memory to track design history and support iterative refinement. It outperforms a strong baseline in producing more complete, modular, and traceable architectures across case studies and real-world specifications.

• MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation 2026-05-30 conf:98% MemGraphRAG introduces a shared-memory multi-agent framework for building higher-quality knowledge graphs in GraphRAG. The shared global context helps agents resolve conflicts and preserve graph consistency, while a memory-aware hierarchical retriever improves performance on complex QA benchmarks.

• Chunking Methods on Retrieval-Augmented Generation - Effectiveness Evaluation Against Computational Cost and Limitations 2026-05-30 conf:97% This paper systematically compares chunking methods for Retrieval-Augmented Generation, analyzing their effectiveness, computational cost, and limitations across scenarios. It shows chunking is a critical design choice in LLM memory retrieval rather than just a simple preprocessing step.

• Eywa: Provenance-Grounded Long-Term Memory for AI Agents 2026-05-29 conf:99% Eywa introduces a provenance-grounded long-term memory system for AI agents that keeps source evidence separate from derived facts and uses deterministic retrieval without LLM calls. It improves auditability and diagnosis of memory failures and reports strong results on LoCoMo, LongMemEval-S, and a technical-memory benchmark.

• ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents 2026-05-29 conf:99% ExpGraph gives LLM agents an external graph-based memory that stores reusable skills and failure lessons from past trajectories. It retrieves and updates these experiences online, helping frozen agents reuse prior knowledge across tasks and improve performance without fine-tuning.

• Beyond Static Dialogues: Benchmarking Realistic, Heterogeneous, and Evolving Long-Term Memory 2026-05-29 conf:98% Introduces RHELM, a benchmark for realistic, evolving long-term memory in LLMs using dynamic dialogues and synchronized heterogeneous sources like documents and emails. It evaluates full-context, RAG, and memory-based systems, showing current methods struggle with multi-source aggregation and contextual reasoning.

• Understanding the Fundamental Design Decisions of Retrieval-Augmented Generation Systems 2026-05-29 conf:98% This paper systematically studies key RAG design choices: when to use RAG, how many documents to retrieve, and how to integrate retrieved information. It shows that effective RAG depends strongly on task and model characteristics rather than one-size-fits-all settings.

• On the impact of retrieved content representations in RAG Pipelines 2026-05-29 conf:97% This paper compares fourteen ways of representing retrieved documents in RAG while keeping retrieval fixed. It shows that preserving answer-bearing content is the main factor driving QA accuracy, while wording, structure, and length matter much less when retention is high.

• Learning Whom to Trust: Market-Feedback Adaptive Retrieval for Frozen LLMs in Event-Driven Financial RAG 2026-05-29 conf:94% The paper adds an external Bayesian source memory to a frozen LLM-based financial RAG system, letting retrieval adapt from market feedback instead of only using text relevance. This improves event-driven prediction and portfolio performance, showing that learning where to retrieve from can outperform tuning the reader.

• Chunking German Legal Code 2026-05-29 conf:91% This paper compares multiple chunking methods for legal RAG on German statutory law and finds that chunks aligned with legal structure, like sections and subsections, outperform more complex LLM-heavy methods. It shows that preserving domain structure improves retrieval quality while reducing computational cost.

• WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction 2026-05-28 conf:99% WorldMemArena is a benchmark for multimodal agent memory that evaluates writing, updating, retrieval, and use of memory during action-world interaction. It compares long-context, RAG, external-memory, and harness-based agents, showing that stronger storage alone does not ensure better performance.

• Less Is More: Elevating RAG via Performance-Driven Context Compression 2026-05-28 conf:98% The paper proposes CORE-RAG, a performance-driven context compression method for RAG that learns which retrieved content to keep using task feedback rather than fixed heuristics. It preserves or improves accuracy while greatly reducing input length and cost.

• RAISE: RAG Design as an Architecture Search Problem 2026-05-28 conf:97% RAISE frames RAG pipeline design as an architecture search problem and provides a benchmark framework for reproducible hyperparameter optimization. It compares 13 search methods across seven datasets, showing that effective RAG configurations are highly task-dependent.

• Bridge-RAG: An Abstract Bridge Tree Based Retrieval Augmented Generation Algorithm 2026-05-28 conf:96% Bridge-RAG improves LLM retrieval by organizing document abstracts into a hierarchical bridge tree that better links queries to relevant chunks. It also adds a Cuckoo Filter for fast entity lookup, boosting both answer quality and retrieval speed.

• CanLegalRAGBench: Evaluating Retrieval-Augmented Generation on Canadian Case Law 2026-05-28 conf:95% CanLegalRAGBench is a benchmark for evaluating legal RAG systems on realistic Canadian case-law questions with expert-annotated answers. It shows that retrieval design choices matter, open-source embeddings can be competitive, and current legal RAG systems still suffer from unsupported claims and evaluation shortcomings.

• SURGENT: A Surgical Multi-Agent Assistance System Across the Perioperative Workflow 2026-05-28 conf:95% SURGENT is a surgical multi-agent LLM system that combines retrieval-augmented reasoning with a memory design for long-term patient history and short-term working summaries. This memory-aware architecture improves consistency and patient-specific perioperative decision support across five clinical tasks.

• Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding 2026-05-28 conf:92% MiA-RAG improves long-context understanding by building a global “mindscape” via hierarchical summarization and using it to guide both retrieval and generation. This helps LLMs connect local evidence with an overall document-level view, outperforming standard RAG baselines.

• Enhancing LLM Medical Coding with Structured External Knowledge 2026-05-28 conf:91% RAG-Coding improves LLM-based medical coding by adding structured external knowledge: a knowledge graph of ICD code relations and concise guideline summaries. It also introduces MDACE-2025, showing better performance and stronger robustness to updated coding guidelines.

• Dial HEALTHDIAL for Advice: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking 2026-05-28 conf:84% HEALTHDIAL is a multilingual spoken dialogue dataset for building and evaluating knowledge-grounded, RAG-based dialogue systems across Arabic, Chinese, English, and Spanish. It also provides benchmarks, a prototype system, and tools for data collection and evaluation.

• MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems 2026-05-27 conf:99% MemTrace introduces a framework to trace how information evolves through LLM memory systems and identify the root causes of memory failures. It also provides a benchmark across systems like long-context and RAG, and uses attribution signals to automatically improve prompts and task performance.

• HGMEM: Hypergraph-based Working Memory to Improve Multi-step RAG for Long-Context Complex Relational Modeling 2026-05-27 conf:99% HGMem introduces a hypergraph-based working memory for multi-step RAG, turning memory from static fact storage into a dynamic structure that captures higher-order relationships. This improves long-context reasoning and global sense-making across challenging benchmarks.

• Beyond Chunk-Local Extraction: Cross-Chunk Graph Augmentation for GraphRAG 2026-05-27 conf:95% CrossAug improves GraphRAG by adding missing cross-chunk relations to its knowledge graph before retrieval. It uses a GNN to find likely gaps and only calls an LLM to complete high-value regions, boosting multi-hop and long-document QA performance.

• SilentRetrieval: Hijacking Retrieval-Augmented Generation via Semantically-Preserving Adversarial Data Poisoning 2026-05-27 conf:94% SilentRetrieval shows how an attacker can poison a RAG corpus with fluent, semantically preserved documents that stay retrievable and steer LLM outputs. It demonstrates strong attack success across datasets, models, and retrievers, and analyzes defense trade-offs.

• P$^2$RAG: Efficient Privacy-Preserving RAG Service Supporting Arbitrary Top-$k$ Retrieval 2026-05-27 conf:93% P²RAG is a privacy-preserving RAG system that supports arbitrary top-k retrieval efficiently without secure sorting. It uses interactive bisection and secret sharing across two non-colluding servers to protect both prompts and databases while substantially speeding up retrieval.

• GraphSteal: Structural Knowledge Stealing from Graph RAG via Traversal Reconstruction 2026-05-27 conf:92% This paper shows that Graph RAG can leak the structure of its hidden knowledge graph through black-box querying. It proposes traversal-based attacks that reconstruct node attributes and graph topology, recovering most of the original graph and exposing privacy risks in memory-augmented LLM systems.

• MM-PoisonRAG: Disrupting Multimodal RAG with Local and Global Poisoning Attacks 2026-05-27 conf:92% This paper introduces MM-PoisonRAG, a framework for attacking multimodal RAG by poisoning external knowledge with targeted or global adversarial content. It shows that even minimal poisoned inputs can mislead retrieval-based generation and defeat existing defenses.

• A Wolf in Sheep's Clothing: Targeted Routing Hijacking in Federated RAG 2026-05-27 conf:90% This paper identifies a new attack on federated RAG where malicious clients manipulate routing profiles to hijack queries and poison retrieved evidence. It proposes a trust-aware post-routing defense that uses evidence feedback to reduce persistent hijacking across routing setups.

• Securing Retrieval-Augmented Generation: A Taxonomy of Attacks, Defenses, and Future Directions 2026-05-27 conf:90% This paper surveys security risks in retrieval-augmented generation by introducing a taxonomy of attacks and defenses across the knowledge-access pipeline. It highlights gaps in current protections and outlines future directions for evaluating and securing RAG, including multimodal and agentic settings.

• Same Question, Different Source, Different Answer: Auditing Source-Dependence in Medical Multi-Source RAG 2026-05-27 conf:89% This paper studies how multi-source RAG systems can give different answers depending on which source is retrieved. It introduces a benchmark, a hierarchical retrieval method, and a judge for measuring source disagreement, improving evaluation of deployed RAG systems.

• CORE: Contrastive Reflection Enables Rapid Improvements in Reasoning 2026-05-27 conf:89% CORE improves LLM reasoning by turning past successful and failed reasoning traces into short natural-language insights that can be reused efficiently. It achieves faster gains than several parametric and memory-based baselines while using fewer rollouts and less context.

• Relevant Is Not Warranted: Evidence-Force Calibration for Cited RAG 2026-05-27 conf:85% This paper introduces FORCEBENCH, a benchmark for testing whether cited RAG systems accurately match claim strength to retrieved evidence. It shows current evaluators often accept overly strong claims and proposes metrics and prompting methods to better assess evidence calibration.

• EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design 2026-05-27 conf:77% EngiAI introduces a multi-agent LLM framework and benchmark suite for engineering design, including tests for working-memory demands and retrieval-augmented parameter selection. It shows how specialized agents coordinated through a supervisor can handle simulation, retrieval, HPC orchestration, and manufacturing workflows.

• MemFail: Stress-Testing Failure Modes of LLM Memory Systems 2026-05-26 conf:99% MemFail is a diagnostic benchmark for stress-testing LLM memory systems. It breaks memory into summarization, storage, and retrieval, builds adversarial datasets for each failure mode, and evaluates modern systems to reveal architectural tradeoffs.

• Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer? 2026-05-26 conf:97% This paper proposes GroundedCache, a safety-focused cache router for RAG that reuses cached answers only when fresh evidence supports them. It introduces gates and a new unsafe-served-rate metric, showing large reductions in wrong cached answers with little latency overhead.

• Detecting Is Not Resolving: The Monitoring Control Gap in Retrieval Augmented LLMs 2026-05-26 conf:97% This paper shows that in multi-turn RAG settings, LLMs can notice contradictory evidence but still fail to act safely on it. It introduces a large-scale evaluation revealing a monitoring-control gap and argues that RAG safety must measure not just detection of conflict, but whether retrieval actually constrains final decisions.

• Query-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion 2026-05-26 conf:93% This paper proposes Query-Adaptive Semantic Chunking, a dynamic method that uses the user query to build more relevant and coherent document chunks for RAG. It outperforms fixed, semantic, and agentic chunking baselines on technical-document retrieval.

• Cordon-MAS: Defending RAG against Knowledge Poisoning via Information-Flow Control 2026-05-26 conf:92% The paper proposes CORDON-MAS, a multi-agent RAG defense that separates evidence extraction, auditing, and answer synthesis using asymmetric memory access. This information-flow design prevents poisoned retrieved documents from influencing final outputs and greatly reduces attack success.

• Mitigating Provenance-Role Collapse in Long-Term Agents via Typed Memory Representation 2026-05-25 conf:99% The paper introduces MemIR, a typed long-term memory format for LLM agents that separates evidence, retrieval cues, and claims to prevent source-monitoring errors. It improves memory retrieval and answer grounding on agent benchmarks, especially for provenance, temporal reasoning, and fragmented evidence.

• Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents 2026-05-25 conf:99% The paper benchmarks personalized memory for long-horizon LLM agents and introduces PerMemBench, a dataset for evaluating user-specific memory policies. It also proposes session-level storage gating to decide when to store interactions, showing personalization helps but accurate gating is still challenging.

• Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory 2026-05-25 conf:98% The paper argues that long-term agent memory is not just storage and introduces Governed Evolving Memory, a state-based model with operators for ingestion, revision, forgetting, and retrieval. It presents MemState, a prototype showing how memory-centric data management can better support persistent AI agents.

• Iterate Until Retrieved: Factual Nugget Optimization for Discoverable Continual Corrections in Agentic RAG 2026-05-25 conf:97% The paper turns factual feedback for agentic RAG systems into compact knowledge entries called factual nuggets and iteratively rewrites them to ensure they are retrievable. It improves how production agents store and use continual corrections across real B2B support settings.

• Retrieval as Reasoning: Self-Evolving Agent-Native Retrieval via LLM-Wiki 2026-05-25 conf:95% The paper introduces LLM-Wiki, an agent-native retrieval system that turns external knowledge into linked wiki-style structures agents can search, read, and traverse as reasoning steps. It also adds persistent self-correction via an Error Book and improves multi-hop QA over strong RAG baselines.

• From PDF to RAG-Ready: Evaluating Document Conversion Frameworks for Domain-Specific Question Answering 2026-05-25 conf:93% The paper evaluates PDF-to-Markdown conversion pipelines for RAG-based question answering and shows that preprocessing quality, especially hierarchy-aware chunking and metadata enrichment, matters more than the conversion tool itself. On a Portuguese administrative-document benchmark, the best pipeline outperformed both a naive baseline and manual curation.

• Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets 2026-05-24 conf:96% This paper shows that compressing tool schemas can free enough context space for agentic RAG to work under tight token budgets. Across many models and settings, schema compression restores or improves retrieval performance when uncompressed tools would overflow the context window.

• H$^{2}$MT: Semantic Hierarchy-Aware Hierarchical Memory Transformer 2026-05-24 conf:96% H²MT organizes long documents into a semantic hierarchy with memory embeddings and routes queries coarse-to-fine to skip irrelevant content. This improves long-context QA efficiency, reducing GPU memory use and latency while maintaining competitive accuracy against RAG and prompt compression baselines.

• M$^\star$: Every Task Deserves Its Own Memory Harness 2026-05-23 conf:99% M* represents an LLM agent’s memory as an executable Python program and automatically evolves task-specific memory designs. Across conversation, planning, and reasoning benchmarks, it outperforms fixed memory systems by jointly optimizing memory schema, storage logic, and workflow instructions.

• The Model Is Not the Product: A Dual-Pillar Architecture for Local-First Psychological Coaching 2026-05-23 conf:98% The paper presents a local-first coaching app that gives an on-device LLM persistent cross-session memory using structured memory cards and semantic/vector retrieval. It emphasizes architecture, orchestration, and benchmarking to achieve reliable long-context behavior on mobile devices.

• Structure-Aware RAG: Structured Retrieval Augmented Generation from Noisy Data for Conversational Agents 2026-05-23 conf:96% The paper proposes SA-RAG, a structured retrieval-augmented generation method that converts noisy retrieved knowledge into tables to reduce irrelevant context for conversational agents. It improves table quality with metadata generation, validation, and optimization, leading to better performance than existing RAG baselines.

• RAGe: A Retrieval-Augmented Generation Evaluation Framework 2026-05-23 conf:95% RAGe is a modular evaluation framework for retrieval-augmented generation systems that benchmarks component choices using accuracy, efficiency, scalability, and hardware telemetry. It helps developers select domain-specific RAG pipelines and prototype them efficiently, even on limited hardware.

• FD-RAG: Federated Dual-System Retrieval-Augmented Generation 2026-05-22 conf:98% FD-RAG is a federated RAG system that separates lightweight memory lookup from on-demand LLM reasoning for edge devices. It builds compact QA memories from local corpora, shares anonymized memories across devices, and improves QA accuracy while greatly reducing latency.

• Evaluating Memory Structure in LLM Agents 2026-05-22 conf:98% The paper introduces StructMemEval, a benchmark for testing whether LLM agents can organize long-term memory into useful structures like ledgers, lists, and trees. It shows simple RAG systems struggle on these tasks, while memory-equipped agents perform better when guided to structure their memory.

• Fine-grained Claim-level RAG Benchmark for Law 2026-05-22 conf:95% Introduces ClaimRAG-LAW, a bilingual legal RAG benchmark with claim-level evaluation for both expert and non-expert users. It enables fine-grained analysis of retrieval and generation errors in legal question answering.

• Asking For An Old Friend: Diagnosing and Mitigating Temporal Failure Modes in LLM-based Statutory Question Answering 2026-05-22 conf:93% This paper studies how LLMs fail on time-sensitive legal questions when relying on stale parametric knowledge. It introduces a benchmark for German statutory QA and shows that temporally filtered RAG substantially improves accuracy over vanilla prompting and web search.

• Through the Stealth Lens: Attention-Aware Defenses Against Poisoning in RAG 2026-05-22 conf:88% This paper studies poisoning attacks in RAG systems and proposes attention-based defenses to detect malicious retrieved passages. It introduces NPAS and an Attention-Variance Filter, improving robustness and showing the limits of stealthy poisoning attacks.

• FATHOMS-RAG: A Framework for the Assessment of Thinking and Observation in Multimodal Systems that use Retrieval Augmented Generation 2026-05-21 conf:95% This paper introduces FATHOMS-RAG, a benchmark for evaluating multimodal RAG pipelines on ingesting, retrieving, and reasoning over text, tables, and images. It also proposes correctness and hallucination metrics and compares open- and closed-source RAG systems.

• RAGCap-Bench: Benchmarking Capabilities of LLMs in Agentic Retrieval Augmented Generation Systems 2026-05-21 conf:94% RAGCap-Bench is a benchmark for measuring the intermediate capabilities LLMs need in agentic retrieval-augmented generation, such as planning, retrieval, and reasoning. It provides fine-grained evaluation of errors and shows that stronger performance on these capabilities correlates with better end-to-end RAG results.

• A Comparative Study of Language Models for Khmer Retrieval-Augmented Question Answering 2026-05-21 conf:93% This paper builds and evaluates a Khmer-language RAG question-answering system for telecom documents. It compares retrievers and generator LLMs, showing retrieval quality is the main bottleneck and that different models trade off grounding, factual accuracy, and answer quality.

• LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems 2026-05-18 conf:99% LongMINT is a benchmark for testing how memory-augmented agents and long-context LLMs handle interference-heavy, evolving information over very long horizons. It shows current systems perform poorly, with retrieval and memory construction as key bottlenecks.

• Long Context Modeling with Ranked Memory-Augmented Retrieval 2026-05-18 conf:99% ERMAR is a memory-augmented retrieval framework for long-context language modeling that dynamically ranks memory entries using relevance scoring and pointwise re-ranking. It leverages historical usage patterns for adaptive retrieval and reports state-of-the-art performance on long-context benchmarks.

• From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG 2026-05-18 conf:98% EPIC builds a compact, preference-aligned memory for on-device RAG by keeping only user-relevant context and steering retrieval toward those preferences. It greatly reduces memory and latency while improving preference-following across multiple tasks.

• Context Memorization for Efficient Long Context Generation 2026-05-18 conf:98% The paper introduces attention-state memory, a training-free external memory that stores precomputed attention states for long prefixes. It improves long-context generation by preserving prefix influence while reducing attention cost, outperforming standard in-context learning and full-attention RAG at lower memory budgets.

• MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning 2026-05-18 conf:98% MemOCR proposes a multimodal memory system that stores interaction history as structured rich text rendered into images, letting agents pack important information more densely than plain text. Trained with budget-aware reinforcement learning, it improves long-horizon reasoning under tight context limits.

• Predictive Prefetching for Retrieval-Augmented Generation 2026-05-18 conf:97% The paper proposes a predictive prefetching framework for RAG that anticipates when retrieval is needed and what to fetch during generation. It reduces latency substantially while preserving answer quality, improving the efficiency of external memory access for LLMs.

• Episodic-Semantic Memory Architecture for Long-Horizon Scientific Agents 2026-05-17 conf:99% The paper proposes a dual-process episodic-semantic memory system for scientific LLM agents that separates a small short-term window from consolidated long-term memory. It shows sustained accuracy and low latency over 15,000-message workflows, outperforming full-context approaches and revealing trade-offs with RAG.

• ConflictRAG: Detecting and Resolving Knowledge Conflicts in Retrieval Augmented Generation 2026-05-17 conf:97% ConflictRAG improves RAG by detecting and resolving conflicts among retrieved documents before generation. It combines efficient conflict detection, source credibility scoring, and a diagnostic metric, leading to more accurate answers across benchmarks and LLM backbones.

• NaviRAG: Towards Active Knowledge Navigation for Retrieval-Augmented Generation 2026-05-17 conf:96% NaviRAG replaces flat RAG retrieval with hierarchical knowledge navigation, letting an LLM agent iteratively retrieve information at the right level of detail. This improves evidence localization, retrieval recall, and long-document QA performance over standard RAG baselines.

• PriHA: A RAG-Enhanced LLM Framework for Primary Healthcare Assistant in Hong Kong 2026-05-17 conf:93% PriHA is a primary healthcare assistant for Hong Kong built on a RAG-enhanced LLM pipeline. It introduces query optimization and a dual-retrieval architecture to combine mixed knowledge sources and improve answer accuracy, clarity, and traceability.

• SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning 2026-05-16 conf:95% SEMA-RAG is a multi-agent medical QA framework that replaces static single-shot retrieval with iterative, role-specialized RAG. It uses interpreter, explorer, and arbiter agents to refine queries, assess retrieval sufficiency, and select evidence-backed answers, improving accuracy across benchmarks and LLMs.

• Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration 2026-05-15 conf:99% This paper studies security risks in LLM agent memory, showing how attackers can plant dormant long-term memories that later trigger data exfiltration. It introduces a benchmark and analysis framework to evaluate attacks and defenses across multiple memory architectures, including RAG and agent memory.

• Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict 2026-05-15 conf:96% This paper studies when RAG follows retrieved context even if that context is wrong. It introduces Context-Driven Decomposition to probe and intervene on context conflict, and releases evaluation settings showing improved robustness to adversarial, temporal, and distractor retrieval errors.

• Don't Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG 2026-05-15 conf:90% The paper proposes Corpus2Skill, which compiles a document corpus into a hierarchical directory of summaries and skill files that an LLM agent can navigate at inference time. Instead of passively consuming retrieved passages, the agent explores the corpus structure, backtracks, and drills down to documents, improving enterprise QA over dense retrieval, RAPTOR, and agentic RAG baselines.

• RAR: Retrieving And Ranking Augmented MLLMs for Visual Recognition 2026-05-15 conf:88% RAR builds an explicit multimodal memory of category examples using CLIP, then retrieves top-k candidates for an MLLM to rank during visual recognition. This retrieval-and-ranking design extends effective context beyond the window limit and improves zero-shot and few-shot fine-grained recognition.

• Deepchecks: Evaluating Retrieval-Augmented Generation (RAG) 2026-05-14 conf:96% Deepchecks presents a framework for evaluating RAG applications through multi-faceted assessment, root-cause analysis, and production monitoring. It aims to measure reliability, relevance, and user satisfaction in systems that use retrieval as external memory for LLMs.

• CuSearch: Curriculum Rollout Sampling via Search Depth for Agentic RAG 2026-05-14 conf:88% CuSearch improves agentic RAG training by prioritizing deeper search trajectories during reinforcement learning. This curriculum-based sampling gives denser supervision to retrieval decisions and boosts answer accuracy across multiple retrieval setups.

• Is Grep All You Need? How Agent Harnesses Reshape Agentic Search 2026-05-14 conf:87% This paper compares grep versus vector retrieval in LLM agent search systems, showing grep often performs better on LongMemEval. It also finds that agent harness design and tool-output format strongly affect retrieval performance, especially with distracting context.

• GAAMA: Graph Augmented Associative Memory for Agents 2026-05-13 conf:99% GAAMA is a long-term memory system for agents that stores episodes, facts, reflections, and concepts in a structured graph to improve retrieval over flat RAG. It combines similarity search, graph-based relevance propagation, and a graph-repair step, outperforming RAG and full-context baselines on multi-session memory benchmarks.

• Thinking Ahead: Prospection-Guided Retrieval of Memory with Language Models 2026-05-13 conf:98% The paper proposes Prospection-Guided Retrieval, which uses imagined future steps as retrieval probes to find user memories that standard similarity-based RAG misses. It also introduces MemoryQuest, a benchmark for long-horizon personalized memory retrieval, and shows large gains in recall and response quality.

• Why Retrieval-Augmented Generation Fails: A Graph Perspective 2026-05-13 conf:97% This paper explains RAG failures by tracing how retrieved evidence flows through transformer layers during answer generation. It finds structural patterns that distinguish successful from failed evidence use, then uses these graphs for error detection and targeted interventions that improve evidence grounding.

• LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues 2026-05-12 conf:99% LongMemEval-V2 is a benchmark for testing long-term memory in web agents, measuring how well memory systems capture environment-specific experience. It evaluates two memory designs, including a RAG-based method and a coding-agent approach, and shows strong gains from explicit memory despite latency tradeoffs.

• Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM Systems 2026-05-12 conf:99% Goal-Mem improves RAG-based memory for conversational LLM agents by using backward-chaining reasoning to break user goals into subgoals and retrieve the right memories for each step. It formalizes this process in natural language logic and shows better performance on multi-hop and implicit reasoning tasks.

• Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation 2026-05-12 conf:99% The paper proposes xMemory, a hierarchical agent memory system that decouples interaction history into reusable components before aggregating them for retrieval. This reduces redundant context and improves answer quality and token efficiency over standard RAG on agent-memory benchmarks.

• Differentially Private Synthetic Text Generation for Retrieval-Augmented Generation (RAG) 2026-05-12 conf:93% The paper proposes DP-SynRAG, a method that builds a differentially private synthetic retrieval database for RAG so it can be reused without repeated privacy loss. It improves privacy-preserving RAG performance under a fixed privacy budget.

• MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models 2026-05-12 conf:91% MoshiRAG adds asynchronous retrieval to full-duplex speech language models so they can fetch external information while keeping conversation natural and real-time. It detects when knowledge is needed, retrieves supporting content during brief conversational gaps, and improves factuality without retraining the base interface model.

• Courtroom-Style Multi-Agent Debate with Progressive RAG and Role-Switching for Controversial Claim Verification 2026-05-12 conf:84% PROClaim frames claim verification as a courtroom-style multi-agent debate with Progressive RAG that iteratively expands evidence during deliberation. It improves zero-shot verification accuracy by combining dynamic retrieval, role-based debate, self-reflection, and multi-judge aggregation.

• ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation 2026-05-11 conf:98% ArchRAG is a graph-based RAG method that builds hierarchical attributed-community indexes to retrieve more relevant knowledge for LLM question answering. It improves answer accuracy while reducing token cost compared with prior graph-based RAG approaches.

• Understand and Accelerate Memory Processing Pipeline for Disaggregated LLM Inference 2026-05-11 conf:96% The paper formalizes LLM memory handling as a four-stage pipeline spanning preparation, relevance computation, retrieval, and inference application. It shows these memory operations are major inference bottlenecks and accelerates them with a GPU-FPGA heterogeneous system, improving speed and energy efficiency.

• ClinicalBench: Stress-Testing Assertion-Aware Retrieval for Cross-Admission Clinical QA on MIMIC-IV 2026-05-11 conf:92% ClinicalBench evaluates assertion-sensitive retrieval for LLM clinical QA and introduces EpiKG, an intent-aware knowledge-graph RAG system that tracks negation, temporality, and attribution. It shows significant gains over dense RAG baselines and releases a benchmark, adjudication data, and tooling.

• Byte-Exact Deduplication in Retrieval-Augmented Generation: A Three-Regime Empirical Analysis Across Public Benchmarks 2026-05-10 conf:96% This paper studies byte-exact deduplication in RAG pipelines as a simple memory/context management method. It shows large context and compute savings in redundant settings while preserving answer quality across multiple model APIs and benchmarks.

• Assessment of RAG and Fine-Tuning for Industrial Question-Answering-Applications 2026-05-10 conf:93% This paper compares RAG and fine-tuning for industrial QA on automotive datasets, jointly measuring answer quality, generation cost, and user interaction cost. It finds that RAG is the most effective and cost-efficient way to adapt both open-source and closed-source LLMs to domain knowledge.

• Generating Leakage-Free Benchmarks for Robust RAG Evaluation 2026-05-09 conf:97% The paper introduces SeedRG, a pipeline for generating RAG benchmarks that avoid knowledge leakage from LLM parametric memory. It creates new retrieval-dependent examples from seed datasets while preserving reasoning structure and filtering out questions answerable without retrieval.

• Human-Inspired Memory Architecture for LLM Agents 2026-05-08 conf:99% This paper introduces a biologically inspired persistent memory system for LLM agents with consolidation, forgetting, reconsolidation, knowledge graphs, and hybrid retrieval. It shows the system can reduce storage while maintaining retrieval quality and improving recall on long-horizon agent benchmarks.

• Defense effectiveness across architectural layers: a mechanistic evaluation of persistent memory attacks on stateful LLM agents 2026-05-08 conf:99% The paper systematically evaluates defenses against persistent memory attacks in stateful LLM agents across architectural layers. It shows most input and retrieval defenses fail, while a memory-layer tool-gating defense largely stops attacks by blocking the recall mechanism they depend on.

• Belief Memory: Agent Memory Under Partial Observability 2026-05-08 conf:99% BeliefMem gives LLM agents a probabilistic external memory that keeps multiple possible conclusions from ambiguous observations instead of committing to one. It updates their probabilities over time and improves agent performance on LoCoMo and ALFWorld in partially observable settings.

• From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms 2026-05-07 conf:99% A survey of LLM agent memory mechanisms that organizes their evolution into Storage, Reflection, and Experience stages. It unifies engineering and cognitive perspectives and highlights design drivers like long-range consistency, dynamic adaptation, and continual learning.

• MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents 2026-05-07 conf:97% The paper studies security risks in persistent external memory for retrieval-augmented LLM agents and introduces MEMSAD, a calibrated anomaly detector for memory poisoning. It provides theoretical guarantees, evaluation corrections, and experiments showing strong detection of continuous attacks while identifying a synonym-based evasion gap.

• Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios 2026-05-07 conf:95% The paper proposes Event-Causal RAG, a memory framework for long-video reasoning that stores semantically segmented events as structured graphs in a dual-store memory. It retrieves causal event chains and supporting video evidence to improve long-range reasoning, memory efficiency, and streaming performance.

• Architecture Matters: Comparing RAG Systems under Knowledge Base Poisoning 2026-05-07 conf:95% This paper compares several RAG architectures under knowledge-base poisoning and shows that system design strongly affects robustness. It finds that failures mostly arise during reasoning over retrieved conflicting content, and introduces a behavioral taxonomy for analyzing these failure modes.

• AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases 2026-05-07 conf:94% AgenticRAG adds tool-based, iterative retrieval to enterprise RAG, letting an LLM search, open, navigate, and summarize documents autonomously. This agentic memory-access design improves recall, factuality, and answer correctness over standard single-shot retrieval pipelines.

• An Agent-Oriented Pluggable Experience-RAG Skill for Experience-Driven Retrieval Strategy Orchestration 2026-05-07 conf:94% The paper introduces Experience-RAG Skill, a pluggable agent module that uses experience memory to choose retrieval strategies for different tasks. It improves retrieval quality across QA, multi-hop reasoning, and scientific verification while keeping the orchestration reusable and modular.

• MiA-Signature: Approximating Global Activation for Long-Context Understanding 2026-05-07 conf:93% The paper introduces MiA-Signature, a compact representation that approximates globally activated context for long-context LLM reasoning. It improves RAG and agentic systems by using selected high-level concepts and lightweight working-memory updates to condition generation more efficiently.

• LeakDojo: Decoding the Leakage Threats of RAG Systems 2026-05-07 conf:93% LeakDojo is a configurable framework for evaluating information leakage in RAG-based LLM systems. It benchmarks multiple attacks across models, datasets, and RAG setups, showing how query generation, instruction-following strength, and faithfulness can increase leakage risk.

• Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall 2026-05-06 conf:99% The paper introduces True Memory, a retrieval-centered agent memory architecture that stores events verbatim and recalls them through a multi-stage retrieval pipeline instead of extraction at ingestion. It achieves strong results on LoCoMo, LongMemEval, and BEAM-1M while running on a single SQLite file without external vector databases or GPUs.

• ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback 2026-05-06 conf:98% ROZA Graphs add persistent reasoning and retrieval graphs to RAG systems so LLM agents can reuse evidence-specific past judgments across runs. This external memory improves answer accuracy, consistency, and efficiency without changing model weights.

• ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical Forgetting 2026-05-05 conf:99% ScrapMem is an on-device long-term memory system for multimodal LLM agents that compresses older memories through optical forgetting and organizes events with an episodic memory graph. It improves storage efficiency and recall while achieving state-of-the-art results on ATM-Bench.

• RAG over Thinking Traces Can Improve Reasoning Tasks 2026-05-05 conf:97% The paper shows that RAG can help reasoning tasks when the retrieval corpus is past thinking traces instead of documents. It introduces T3, a method to structure these traces for retrieval, improving math, coding, and science reasoning while sometimes lowering inference cost.

• Do Multimodal RAG Systems Leak Data? A Comprehensive Evaluation of Membership Inference and Image Caption Retrieval Attacks 2026-05-05 conf:85% This paper evaluates privacy leakage in multimodal RAG systems by testing whether attackers can infer if an image is in the retrieval corpus and recover related captions. It shows that connecting private visual datasets through mRAG can expose sensitive information.

• MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing 2026-05-04 conf:99% MEMAUDIT introduces an exact evaluation protocol for long-term LLM memory writing under storage budgets. It isolates the quality of what memory systems preserve from downstream retrieval and QA effects using auditable optimization packages and certified solvers.

• Learning from Supervision with Semantic and Episodic Memory: A Reflective Approach to Agent Adaptation 2026-05-04 conf:99% The paper proposes a memory-augmented LLM agent that stores critique-based episodic memories and distills them into semantic guidance to learn from labeled examples without fine-tuning. It improves task accuracy over zero-shot and label-only RAG baselines while reducing reasoning token costs.

• AdaGATE: Adaptive Gap-Aware Token-Efficient Evidence Assembly for Multi-Hop Retrieval-Augmented Generation 2026-05-04 conf:98% AdaGATE is a training-free controller for multi-hop RAG that repairs missing evidence under a token budget using gap tracking, micro-queries, and utility-based selection. It improves evidence quality and robustness to noisy or redundant retrieval while using far fewer input tokens.

• Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction 2026-05-04 conf:96% This paper proposes a two-stage method to make RAG more dependable by filtering retrieved chunks with conformal prediction and estimating whether generated answers stay faithful to the retrieved context. It improves answer quality and provides confidence signals for factual consistency.

• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study 2026-05-04 conf:95% This paper systematically compares five retrieval strategies in a biomedical RAG pipeline for LLM question answering. It shows cross-encoder reranking performs best overall and provides a controlled evaluation framework, metrics, and open-source code for assessing RAG retrieval quality.

• Retrieval-Augmented LLMs for Security Incident Analysis 2026-05-04 conf:90% The paper presents a RAG-based system for security incident analysis that filters logs with targeted queries and retrieves relevant evidence for LLM reasoning. It shows accurate, cost-effective attack reconstruction across multiple LLMs, with retrieval proving essential for handling large forensic contexts.

• When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in Scientific Multi-hop Question Answering 2026-05-03 conf:98% This paper diagnoses when iterative RAG helps LLMs more than giving all evidence at once in scientific multi-hop QA. It shows staged retrieval and reasoning can outperform oracle static context by reducing overload, correcting early mistakes, and improving retrieval-control behavior.

• AutoRAGTuner: A Declarative Framework for Automatic Optimization of RAG Pipelines 2026-05-03 conf:96% AutoRAGTuner is a declarative framework that automates building, evaluating, and tuning RAG pipelines. It uses modular components, a unified data model, and Bayesian optimization to improve performance across vanilla and graph-based RAG while greatly reducing engineering effort.

• Improving Factuality in LLMs via Inference-Time Knowledge Graph Construction 2026-05-03 conf:88% The paper proposes building and refining a knowledge graph during inference by combining LLM-extracted knowledge with external retrieval. This structured retrieval approach improves factual QA accuracy over standard baselines and offers a more interpretable alternative to text-only RAG.

• ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems 2026-05-02 conf:99% ZenBrain proposes a neuroscience-inspired 7-layer memory architecture for autonomous AI systems that integrates 15 memory mechanisms under a unified coordinator. It improves long-context and answer-quality performance while reducing token cost, storage, and instability compared with existing memory systems.

• Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation 2026-05-02 conf:95% This paper proposes CoRM-RAG, a retrieval framework that prioritizes evidence robustness over semantic similarity in RAG. It trains an evidence critic using counterfactual query perturbations to retrieve documents that help LLMs stay correct under biased or adversarial prompts.

• Verbal-R3: Verbal Reranker as the Missing Bridge between Retrieval and Reasoning 2026-05-02 conf:89% Verbal-R3 improves RAG by adding a verbal reranker that explains why retrieved passages matter, helping an LLM connect retrieval results to its reasoning. Its agentic generator-reranker loop achieves stronger performance on complex question answering.

• H-RAG at SemEval-2026 Task 8: Hierarchical Parent-Child Retrieval for Multi-Turn RAG Conversations 2026-05-01 conf:96% The paper proposes a hierarchical RAG pipeline for multi-turn conversations that retrieves fine-grained child chunks and reconstructs broader parent-level context for generation. It shows that parent-level evidence aggregation and hybrid retrieval improve conversational RAG performance.

• "I Don't Know" -- Towards Appropriate Trust with Certainty-Aware Retrieval Augmented Generation 2026-05-01 conf:95% The paper introduces CERTA, a certainty-aware RAG system that estimates uncertainty from question, context, and answer relevance to improve trustworthiness. It also presents a benchmark for non-objective questions, showing more cautious and less overconfident behavior than baseline RAG.

• Lightweight Domain Adaptation of a Large Language Model for Legal Assistance in the Indian Context 2026-05-01 conf:92% The paper presents a legal-assistance system for India that combines an 8B Llama model with RAG over a curated legal corpus. It shows strong benchmark performance, reduced hallucinations, and higher parameter efficiency than a much larger baseline.

• E-MIA: Exam-Style Black-Box Membership Inference Attacks against RAG Systems 2026-05-01 conf:91% This paper presents E-MIA, a black-box membership inference attack against RAG systems that turns facts from a candidate document into exam-style questions. By grading the model’s answers, it infers whether the document is in the retrieval corpus more accurately and stealthily than prior methods.

• When RAG Chatbots Expose Their Backend: An Anonymized Case Study of Privacy and Security Risks in Patient-Facing Medical AI 2026-05-01 conf:84% This case study shows that a medical RAG chatbot exposed its system prompt, retrieval setup, knowledge-base contents, and recent patient conversations through browser-accessible backend traffic. It highlights major privacy and governance risks in deploying memory-enabled medical LLM systems.

• NeocorRAG: Less Irrelevant Information, More Explicit Evidence, and More Effective Recall via Evidence Chains 2026-04-30 conf:97% NeocorRAG improves RAG by optimizing retrieval quality through evidence chains, balancing recall with more useful evidence for reasoning. It introduces the Recall Conversion Rate metric and a training-free framework that achieves stronger QA performance with far fewer tokens.

• Iterative Multimodal Retrieval-Augmented Generation for Medical Question Answering 2026-04-30 conf:97% MED-VRAG is an iterative multimodal RAG system for medical QA that retrieves page images instead of text chunks and uses a memory bank to accumulate evidence across reasoning rounds. It improves accuracy over no-retrieval and text-only retrieval baselines, showing gains from both iteration and explicit memory.

• When to Retrieve During Reasoning: Adaptive Retrieval for Large Reasoning Models 2026-04-29 conf:98% The paper introduces ReaLM-Retrieve, a framework that decides when to retrieve external evidence during multi-step reasoning instead of only before generation. It improves answer accuracy and retrieval efficiency by detecting step-level knowledge gaps and triggering retrieval only when it is most useful.

• Faithfulness-QA: A Counterfactual Entity Substitution Dataset for Training Context-Faithful RAG Models 2026-04-29 conf:98% This paper introduces Faithfulness-QA, a large-scale counterfactual QA dataset designed to train and evaluate RAG models to rely on retrieved context instead of internal knowledge. It creates controlled conflicts between context and parametric memory using entity substitution.

• LIT-RAGBench: Benchmarking Generator Capabilities of Large Language Models in Retrieval-Augmented Generation 2026-04-29 conf:97% LIT-RAGBench is a benchmark for testing how well LLM generators handle RAG tasks, including evidence integration, reasoning, table understanding, logic, and abstention. It provides Japanese and English datasets plus evaluation code to measure generator strengths and weaknesses in practical RAG settings.

• Auto-ARGUE: LLM-Based Report Generation Evaluation 2026-04-29 conf:79% Auto-ARGUE is an LLM-based evaluation framework for citation-backed report generation in RAG systems. It shows good agreement with human judgments on TREC tasks and includes a visualization tool for detailed analysis.

• Citation Failure: Definition, Analysis and Efficient Mitigation 2026-04-28 conf:89% This paper studies why RAG systems fail to cite complete supporting evidence even when their answers are correct. It introduces a benchmark for analyzing citation failure and a framework that combines generative, attention-based, and retrieval methods to improve citation quality efficiently.

• Retrieval-Augmented LLMs for Evidence Localization in Clinical Trial Recruitment from Longitudinal EHR Narratives 2026-04-28 conf:84% This paper compares long-context prompting, extractive summarization, and retrieval-augmented generation for locating evidence in long EHR narratives for clinical trial recruitment. It finds that RAG with a medical-adapted LLM performs best, improving screening on criteria requiring reasoning across long documents.

• MEMCoder: Multi-dimensional Evolving Memory for Private-Library-Oriented Code Generation 2026-04-27 conf:98% MEMCoder gives LLMs an evolving external memory that stores distilled coding guidelines from past attempts and retrieves them alongside API docs at inference time. Using execution feedback to update and resolve conflicts in memory, it significantly improves private-library code generation over standard RAG.

• In-depth Analysis of Graph-based RAG in a Unified Framework 2026-04-27 conf:97% The paper presents a unified framework for graph-based RAG and systematically compares representative methods under consistent settings across QA tasks. It also identifies improved method variants and highlights future research directions for memory-enhanced LLM retrieval.

• Reducing Redundancy in Retrieval-Augmented Generation through Chunk Filtering 2026-04-27 conf:95% The paper reduces redundancy in RAG indexes by filtering chunks using semantic, topic, and named-entity signals. It shows entity-based filtering can shrink vector indexes by 25–36% while keeping retrieval quality close to the original.

• CUB: Benchmarking Context Utilisation Techniques for Language Models 2026-04-27 conf:95% CUB introduces a benchmark for testing how well language models use retrieved context in RAG under noisy conditions. It systematically compares seven context-utilisation methods across multiple datasets and models, showing that current methods often fail on realistic retrieval scenarios.

• BRIEF-Pro: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning 2026-04-27 conf:94% BRIEF-Pro is a lightweight context compressor for RAG that summarizes long retrieved documents into short query-focused evidence for multi-hop QA. It enables controllable compression, improves answer accuracy over prior compression methods, and reduces computational cost.

• AdaComp: Extractive Context Compression with Adaptive Predictor for Retrieval-Augmented Large Language Models 2026-04-26 conf:97% AdaComp is a low-cost RAG context compression method that predicts how many retrieved documents are actually needed based on query complexity and retrieval quality. It cuts inference cost while keeping QA performance close to using the full retrieved context.

• S2G-RAG: Structured Sufficiency and Gap Judging for Iterative Retrieval-Augmented QA 2026-04-26 conf:96% S2G-RAG improves iterative multi-hop RAG by judging whether current evidence memory is sufficient and identifying missing information to guide the next retrieval step. It also compresses context into sentence-level evidence to reduce noise and improve QA robustness.

• Evolve: A Persistent Knowledge Lifecycle for Small Language Models 2026-04-25 conf:98% Evolve adds a persistent external knowledge store to a small language model, with staged writing, offline consolidation, refresh, and section-based retrieval. It boosts accuracy substantially while reducing expensive teacher-model calls through reusable memory.

• Knowledge Capsules: Structured Nonparametric Memory Units for LLMs 2026-04-23 conf:99% The paper introduces Knowledge Capsules, structured external memory units that store relational knowledge for frozen LLMs. Instead of adding retrieved text to the prompt like RAG, it injects memory as attention-compatible key-value representations so external knowledge can directly influence the model. This improves stability and performance on long-context and multi-hop QA without retraining.

• Spatial Metaphors for LLM Memory: A Critical Analysis of the MemPalace Architecture 2026-04-23 conf:98% This paper critically analyzes MemPalace, an open-source long-term memory system for LLMs, and argues its gains come mainly from verbatim storage and standard vector retrieval rather than its spatial metaphor. It highlights useful design ideas like deterministic zero-LLM writes, low wake-up cost, and a layered memory stack.

• Self-Aware Vector Embeddings for Retrieval-Augmented Generation: A Neuroscience-Inspired Framework for Temporal, Confidence-Weighted, and Relational Knowledge 2026-04-22 conf:98% This paper presents SmartVector, a memory-aware RAG framework that enriches embeddings with time, confidence, and relationship information instead of treating them as static vectors. It adds consolidation and update mechanisms inspired by biological memory and shows better accuracy, calibration, and stale-information handling on versioned retrieval tasks.

• Improving End-to-End Training of Retrieval-Augmented Generation Models via Joint Stochastic Approximation 2026-04-22 conf:95% This paper improves end-to-end training of retrieval-augmented generation models, where retrieved passages serve as external memory for the model. It proposes a joint stochastic approximation method to better optimize retrieval and generation together, reducing gradient estimation problems and improving performance on open-domain QA and knowledge-grounded dialogue.

• Coverage, Not Averages: Semantic Stratification for Trustworthy Retrieval Evaluation 2026-04-22 conf:87% This paper proposes semantic stratification, a way to evaluate retrieval in RAG by organizing a corpus into semantic clusters and generating queries to cover missing areas. The method provides better coverage guarantees and more interpretable analysis of retrieval failures, leading to more reliable retrieval evaluation than standard aggregate metrics.

• HaS: Accelerating RAG through Homology-Aware Speculative Retrieval 2026-04-22 conf:87% This paper presents HaS, a framework that speeds up retrieval-augmented generation by speculatively reusing results from previously seen homologous queries and validating whether they satisfy the new query. It reduces retrieval latency substantially with only a small accuracy drop, and also helps accelerate multi-hop agentic RAG pipelines.

• Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems 2026-04-21 conf:99% This paper presents the Mesh Memory Protocol, a semantic memory infrastructure for multi-agent LLM systems that lets agents share and maintain persistent cognitive state across sessions. It introduces structured memory blocks, field-level acceptance rules, lineage tracking, and remix-based storage so agents can build traceable, role-aware long-term collaborative memory.

• Stable-RAG: Mitigating Retrieval-Permutation-Induced Hallucinations in Retrieval-Augmented Generation 2026-04-21 conf:97% This paper examines how the order of retrieved documents in RAG can change an LLM’s answers, even when the correct document is present. It proposes Stable-RAG, which tests multiple retrieval orders and aggregates internal reasoning patterns to reduce hallucinations and make RAG outputs more consistent and accurate.

• VimRAG: Navigating Massive Visual Context in Retrieval-Augmented Generation via Multimodal Memory Graph 2026-04-21 conf:96% VimRAG is a multimodal RAG framework that organizes agent states and retrieved text/image/video evidence into a dynamic memory graph. It uses graph-aware memory encoding to allocate more visual tokens to important evidence while compressing or removing less useful clues, improving long-context multimodal reasoning.

• Forage V2: Knowledge Evolution and Transfer in Autonomous Agent Organizations 2026-04-21 conf:92% Forage V2 studies organizational memory for autonomous agent systems, where knowledge is stored as transferable readable documents and accumulated across runs. It shows that this memory improves coverage estimation, reduces cost and rounds, and can transfer useful knowledge from stronger to weaker agents.

• ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios 2026-04-21 conf:88% ViDoRe V3 is a multimodal benchmark for evaluating RAG systems in realistic, visually rich document settings. It measures retrieval, grounding, and answer quality across multilingual, multi-document queries, and shows that current RAG pipelines still struggle with non-textual content and fine-grained visual grounding.

• RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora 2026-04-21 conf:86% This paper introduces RARE, a framework for evaluating RAG retrieval on highly redundant, high-similarity corpora such as finance, legal, and patent documents. It tracks atomic facts across overlapping documents and builds more realistic benchmarks, showing that retrievers that look strong on standard datasets can fail badly in real-world redundant settings.

• SpecAgent: A Speculative Retrieval and Forecasting Agent for Code Completion 2026-04-21 conf:80% SpecAgent is a code-completion system that proactively explores a software repository at indexing time to build speculative context for future edits, improving both latency and completion quality. It reframes repository retrieval as asynchronous external memory preparation and introduces a leakage-free benchmark for evaluating repository-aware retrieval methods.

• WorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time Reconciliation 2026-04-20 conf:99% WorldDB is a persistent memory engine for long-running LLM agents that replaces flat vector-store RAG with a recursive graph-of-worlds structure, immutable content-addressed nodes, and programmable edge behaviors for updates, contradictions, and merges. It is explicitly framed as an external memory system and reports strong gains on LongMemEval for recall, temporal reasoning, and knowledge updates.

• Procedural Knowledge at Scale Improves Reasoning 2026-04-20 conf:98% This paper introduces Reasoning Memory, a retrieval-based external memory for reasoning models that reuses procedural knowledge from past reasoning traces. It breaks trajectories into subquestion-subroutine pairs, retrieves relevant procedures during inference, and shows improved reasoning performance across math, science, and coding benchmarks.

• Is Agentic RAG worth it? An experimental comparison of RAG approaches 2026-04-20 conf:94% This paper experimentally compares enhanced RAG pipelines with agentic RAG, where an LLM decides how and when to retrieve or iterate. It evaluates trade-offs in effectiveness and cost across scenarios, offering guidance on which RAG design works best in practice.

• Evaluating Multi-Hop Reasoning in RAG Systems: A Comparison of LLM-Based Retriever Evaluation Strategies 2026-04-20 conf:89% The paper studies how to better evaluate retrieval in RAG systems for multi-hop questions, where multiple pieces of evidence must be combined. It proposes CARE, a context-aware LLM-as-judge method that more accurately assesses retriever quality than prior strategies across several QA benchmarks.

• Semantic Entanglement in Vector-Based Retrieval: A Formal Framework and Context-Conditioned Disentanglement Pipeline for Agentic RAG Systems 2026-04-20 conf:89% This paper studies a failure mode in RAG systems where mixed-topic documents create overlapping embeddings that hurt retrieval precision. It formalizes this as semantic entanglement, introduces an Entanglement Index, and proposes a preprocessing pipeline with agent-performance feedback that substantially improves Top-K retrieval on a real enterprise knowledge base.

• ArbGraph: Conflict-Aware Evidence Arbitration for Reliable Long-Form Retrieval-Augmented Generation 2026-04-20 conf:88% ArbGraph improves long-form retrieval-augmented generation by resolving conflicts among retrieved evidence before generation. It turns documents into atomic claims, builds a graph of support and contradiction relations, and iteratively scores credibility so unreliable claims are filtered out. This leads to more factual and less hallucinated long-form answers on standard RAG benchmarks.

• Beyond Explicit Refusals: Soft-Failure Attacks on Retrieval-Augmented Generation 2026-04-20 conf:87% This paper studies a subtle attack on RAG systems that causes the model to return fluent but low-utility answers instead of obvious refusals or crashes. It introduces a black-box method for crafting adversarial retrieved documents and shows that these attacks reliably degrade answer usefulness across different RAG setups and models.

• Domain-oriented RAG Assessment (DoRA): Synthetic Benchmarking for RAG-based Question Answering on Defense Documents 2026-04-20 conf:78% DoRA is a synthetic benchmark for testing RAG-based question answering on defense documents using auditable evidence passages and multiple intent-conditioned question types. It is designed to reduce contamination effects from pretraining and evaluate task success and faithfulness under domain shift.

• MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization 2026-04-19 conf:99% MetaMem is an LLM memory framework that adds a self-evolving meta-memory to help the model use stored interaction history more effectively. It self-reflects on reasoning processes to distill reusable experience for selecting and integrating evidence from fragmented memories, improving long-horizon interaction performance.

• The Continuity Layer: Why Intelligence Needs an Architecture for What It Carries Forward 2026-04-19 conf:98% This position paper argues that AI systems need a dedicated continuity layer to carry understanding across sessions instead of relying only on context windows or flat memory retrieval. It proposes a storage primitive called Decomposed Trace Convergence Memory, frames continuity as a distinct system property, and connects the idea to benchmark-based evaluation of long-term memory behavior.

• Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning 2026-04-19 conf:95% This paper proposes GRIP, a framework that turns retrieval into part of token generation so the model can decide when to retrieve, how to rewrite queries, and when to stop. It tightly couples retrieval with reasoning for multi-step question answering and outperforms strong RAG baselines on several benchmarks.

• ATANT: An Evaluation Framework for AI Continuity 2026-04-19 conf:95% ATANT is an evaluation framework and benchmark for testing whether AI systems can maintain continuity over time by storing, updating, and correctly retrieving context. It provides a story-based corpus and multi-checkpoint methodology to measure memory-system behavior such as persistence, disambiguation, and avoiding cross-contamination across many coexisting narratives.

• ATANT v1.1: Positioning Continuity Evaluation Against Memory, Long-Context, and Agentic-Memory Benchmarks 2026-04-19 conf:88% This paper analyzes how ATANT’s continuity evaluation differs from existing LLM memory, long-context, and agentic-memory benchmarks. It argues that benchmarks like LOCOMO, LongMemEval, MemoryBench, Zep, and Letta/MemGPT measure related capabilities but do not capture the continuity properties ATANT defines, and it documents benchmark-specific methodological gaps.

• Align Documents to Questions: Question-Oriented Document Rewriting for Retrieval-Augmented Generation 2026-04-19 conf:83% This paper improves retrieval-augmented generation by rewriting retrieved documents into a question-oriented style that LLMs can use more effectively while preserving factual content. It introduces a plug-and-play rewriter for RAG pipelines and shows consistent gains in answer quality with little extra latency.

• LTRR: Learning To Rank Retrievers for LLMs 2026-04-18 conf:93% This paper improves retrieval-augmented generation by learning which retriever to use for each query instead of relying on a single fixed retriever. It ranks candidate retrievers based on how much they help the LLM answer correctly, leading to better QA performance and stronger generalization.

• Illocutionary Explanation Planning for Source-Faithful Explanations in Retrieval-Augmented Language Models 2026-04-18 conf:78% This paper studies retrieval-augmented LLMs for generating explanations grounded in authoritative textbook sources. It benchmarks source faithfulness and proposes chain-of-illocution prompting to improve retrieval-guided, traceable explanations, showing better adherence to sources across several models.

• Evaluating Memory Capability in Continuous Lifelog Scenario 2026-04-17 conf:99% This paper introduces LifeDialBench, a benchmark for testing LLM memory systems in continuous lifelog scenarios, along with an online evaluation protocol that respects temporal order. Its experiments show that current complex memory designs can underperform a simple RAG baseline, suggesting that preserving context faithfully is crucial for this kind of memory task.

• What Deserves Memory: Adaptive Memory Distillation for LLM Agents 2026-04-16 conf:99% The paper proposes NEMORI, a memory distillation framework for LLM agents that learns what experiences are worth retaining instead of relying on hand-designed memory heuristics. It integrates interactions into episodic narratives and distills semantic knowledge based on prediction error, improving efficiency, storage use, and downstream performance.

• Feedback Adaptation for Retrieval-Augmented Generation 2026-04-16 conf:97% This paper studies how RAG systems adapt after users provide corrective feedback. It introduces metrics for how quickly feedback changes future behavior and how well that correction generalizes, and proposes PatchRAG, an inference-time method that applies feedback immediately without retraining.

• Guaranteeing Knowledge Integration with Joint Decoding for Retrieval-Augmented Generation 2026-04-15 conf:92% The paper proposes GuarantRAG, a RAG framework that separates an LLM’s internal answer from an evidence-grounded answer and then combines them with joint decoding. This improves how the model uses retrieved external knowledge, reducing hallucinations and improving QA accuracy.

• Beyond Black-Box Interventions: Latent Probing for Faithful Retrieval-Augmented Generation 2026-04-15 conf:86% The paper proposes ProbeRAG, a retrieval-augmented generation framework that improves faithfulness by probing the model’s latent states to detect conflicts between retrieved context and internal knowledge. It combines context pruning, latent conflict detection, and conflict-aware attention to help the model better use retrieved evidence.

KV Cache & Attention Memory

KV cache compression, eviction, offloading, and efficient attention for long-context inference.

• QCFuse: Query-Aware Cache Fusion via Compressed View for Efficient RAG Serving 2026-06-04 conf:93% QCFuse speeds up RAG serving by using a compressed, query-aware selector to decide which retrieved context tokens need recomputation during KV-cache fusion. Implemented in SGLang, it matches full-prefill quality while reducing prefill time versus full prefill and prior cache-fusion methods.

• RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention 2026-06-04 conf:91% RedKnot redesigns the LLM KV cache as a head-aware memory system instead of a single monolithic buffer. This enables more efficient long-context serving through selective KV reuse, compression, hot/cold separation, and distributed placement without retraining.

• Cartridges at Scale: Training Modular KV Caches over Large Document Collections 2026-06-03 conf:98% The paper proposes Cartridges at Scale, a framework for training modular KV-cache memories over large document collections so LLMs can reuse compressed document representations instead of rereading long contexts. It adds scalable multi-cartridge training and storage management, achieving RAG-level or better accuracy with far fewer prompt tokens.

• LazyAttention: Efficient Retrieval-Augmented Generation with Deferred Positional Encoding 2026-06-03 conf:92% LazyAttention improves retrieval-augmented generation by enabling position-agnostic reuse of KV caches without copying or re-encoding them. This cuts latency and boosts throughput while preserving output quality.

• Value-Aware Stochastic KV Cache Eviction for Reasoning Models 2026-06-02 conf:91% The paper proposes VaSE, a training-free KV cache eviction method for reasoning models that protects high-value states and uses stochastic eviction to avoid failures and improve cache diversity. It achieves better accuracy than prior eviction and selection methods at the same compression level while keeping memory usage static.

• AURA: Action-Gated Memory for Robot Policies at Constant VRAM 2026-06-01 conf:95% AURA adds a constant-size recurrent memory to a frozen vision-language-action model and learns when to write based on whether new observations would change the next action. It keeps memory use fixed while matching baseline performance and greatly reducing memory writes in long-horizon robot tasks.

• Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction 2026-06-01 conf:94% The paper proposes LU-KV, a global optimization framework for KV cache eviction that allocates token budgets across attention heads based on long-horizon utility. It preserves effective short-term memory while cutting KV cache size and inference cost with minimal accuracy loss.

• ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution 2026-06-01 conf:92% ForesightKV learns which KV cache entries to evict during long reasoning so LLMs can handle longer sequences with less memory cost. It combines supervised distillation from future attention-based eviction traces with reinforcement learning, improving reasoning performance under tight cache budgets.

• STaR-KV: Spatio-Temporal Adaptive Re-weighting for KV Cache Compression in GUI Vision-Language Models 2026-06-01 conf:89% STaR-KV is a training-free method for compressing KV cache memory in GUI vision-language agents by adaptively reweighting token importance across spatial subspaces and time. It reduces GPU memory use while preserving or improving task accuracy on GUI benchmarks.

• Multi-Segment Attention: Enabling Efficient KV-Cache Management for Faster Large Language Model Serving 2026-06-01 conf:86% AsymCache improves LLM serving by managing KV-cache placement and eviction based on GPU attention efficiency, not just access patterns. It introduces Multi-Segment Attention, position-aware eviction, and adaptive chunking to speed exact inference while preserving outputs.

• GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs 2026-05-29 conf:94% GRKV is a training-free KV-cache compression method for long-context LLMs that reduces information loss during eviction and merging. It uses ridge regression to better distribute evicted-token information across retained tokens, improving performance on LongBench and RULER with low overhead.

• Model-Native Computing Architecture: Envisioning Future System Architecture Through the Lens of Computer Architecture 2026-05-29 conf:92% A conceptual survey proposing a six-layer model-native computing architecture for LLM systems. It unifies memory management, context budgeting, KV-cache reuse, and agent coordination under explicit system design laws for building scalable LLM-based agents.

• OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference 2026-05-28 conf:95% OBCache improves long-context LLM inference by pruning KV cache entries with output-aware saliency scores instead of heuristic attention-based ranking. It frames cache eviction as structured pruning and shows better long-context accuracy on LLaMA and Qwen.

• ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs 2026-05-28 conf:94% ParisKV is a GPU-native framework for fast, drift-robust KV-cache retrieval in long-context LLMs, including million-token settings with CPU-offloaded caches. It improves decoding efficiency and latency while matching or exceeding full-attention quality.

• xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction 2026-05-27 conf:89% xKV compresses LLM KV-cache by jointly factorizing cache states across layers into a shared low-rank subspace. It cuts long-context inference memory up to 8x and, with selective reconstruction, speeds decoding while largely preserving accuracy.

• Which Heads Matter for Reasoning? RL-Guided KV Cache Compression 2026-05-27 conf:84% The paper proposes RLKV, which uses reinforcement learning to identify attention heads most important for reasoning and allocates KV cache selectively to them. This enables 20–60% cache reduction with near-lossless reasoning performance and faster inference.

• NestedKV: Nested Memory Routing for Long-Context KV Cache Compression 2026-05-26 conf:97% NestedKV compresses LLM KV cache for long-context inference by routing tokens using multi-time-scale memory signals and adaptive per-head budgets. It is training-free, needs no model changes, and improves performance especially when only a small cache can be kept.

• SPHERICAL KV: Angle-Domain Attention and Rate-Distortion Retention for Efficient Long-Context Inference 2026-05-26 conf:94% The paper proposes Spherical KV, a KV-cache compression method for long-context LLM inference that stores keys in angle-based form and optimizes token retention and precision under a memory budget. This reduces memory traffic during decoding while preserving efficiency in realistic serving settings.

• CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM 2026-05-24 conf:98% CONF-KV is a confidence-aware KV-cache manager for long-horizon LLM inference that adapts cache size based on model uncertainty and ranks tokens by attention and recency. It cuts memory substantially while preserving performance on long-context retrieval and agent-style web tasks.

• Polynomial Context-Truncation Sensitivity in Autoregressive Language Models: Sequential Wyner-Ziv Bounds for KV Cache Compression 2026-05-24 conf:96% This paper analyzes KV cache compression in autoregressive language models and shows that sensitivity to truncated context follows a power law. It derives matching bounds for sliding-window cache policies and finds that recency-based retention strongly outperforms random cache eviction.

• Meta-Soft: Leveraging Composable Meta-Tokens for Context-Preserving KV Cache Compression 2026-05-23 conf:96% Meta-Soft compresses LLM KV caches by dynamically generating prompt-specific soft tokens and integrating information from evicted tokens into retained ones. This preserves context better than prior eviction methods while improving long-context decoding efficiency.

• Adaptive Mass-Segmented KV Compression for Long-Context Reasoning 2026-05-22 conf:98% The paper proposes AMS, a region-aware KV-cache compression method for long-context LLMs that preserves important reasoning segments instead of evicting tokens purely by global ranking. It plugs into existing KV compression methods and improves long-context reasoning and serving efficiency.

• Tensor Cache: Eviction-conditioned Associative Memory for Transformers 2026-05-21 conf:98% Tensor Cache adds a fixed-size second-level associative memory to transformers that stores KV pairs evicted from the sliding attention window. This lets models recover useful old context efficiently and improves the tradeoff between bounded memory use and long-context quality.

• CacheClip: Accelerating RAG with Effective KV Cache Reuse 2026-05-21 conf:97% CacheClip speeds up RAG by reusing and selectively recomputing KV caches, reducing time-to-first-token while preserving answer quality. It uses a small auxiliary model to identify important tokens for restoring cross-chunk attention and adds practical system optimizations for deployment.

• LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems 2026-05-21 conf:94% LCGuard makes KV-cache sharing safer in multi-agent LLMs by treating shared caches as latent working memory and transforming them to reduce sensitive information leakage. It uses adversarial training to block reconstruction of private inputs while preserving task performance.

• ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning 2026-05-21 conf:94% ArborKV is a structure-aware KV cache management method for tree-based LLM reasoning that evicts low-value inactive branches and lazily restores them when needed. It cuts peak KV memory by up to 4x while largely preserving reasoning accuracy, enabling deeper and wider search under fixed hardware limits.

• KVDrive: A Holistic Multi-Tier KV Cache Management System for Long-Context LLM Inference 2026-05-18 conf:96% KVDrive is a systems approach for managing LLM KV cache across GPU, DRAM, and SSD to support long-context inference efficiently. It coordinates cache placement, scheduling, and cross-tier data movement to reduce latency and improve throughput while preserving accuracy.

• Protection Is (Nearly) All You Need: Structural Protection Dominates Scoring in Globally Capped KV Eviction 2026-05-18 conf:92% This paper shows that protecting structurally important boundary tokens in the KV cache matters more than sophisticated eviction scoring for long-context decoding. A simple reserved-cache strategy recovers much of full-cache performance at small cache sizes, with scoring refinements providing smaller additional gains.

• Minimal-Intervention KV Retention: A Design-Space Study and a Diversity-Penalty Survivor 2026-05-17 conf:95% This paper compares KV-cache compression strategies for long-context LLM reasoning and finds most tested designs ineffective at small budgets. It proposes a simple diversity-penalized token retention scorer that outperforms heavier redesigns in some settings under a controlled evaluation protocol.

• Key-Value Means: Transformers with Expandable Block-Recurrent Compressed Memory 2026-05-15 conf:95% KVM introduces an expandable compressed memory mechanism for transformers that supports long-context processing with subquadratic prefill and sublinear state growth. It combines benefits of transformers and RNNs while reducing KV-cache usage and remaining easy to implement with standard operations.

• Self-Pruned Key-Value Attention: Learning When to Write by Predicting Future Utility 2026-05-13 conf:97% The paper proposes SP-KV, a method that predicts which past key-value states will be useful later and only keeps those in the long-term KV cache. This dynamically shrinks LLM memory use and speeds decoding with minimal quality loss.

• EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction 2026-05-13 conf:93% EchoKV compresses LLM KV caches for long-context inference while allowing flexible switching back to full-cache mode. It reconstructs discarded KV components from partial caches using similarity across attention heads and achieves better long-context performance with minimal fine-tuning cost.

• KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference 2026-05-12 conf:97% KV-Fold introduces a training-free way to extend LLM context by recursively carrying forward and appending the KV cache across chunks. It enables stable long-context inference and exact long-range retrieval up to 128K tokens without retraining or architectural changes.

• GRC: Unifying Reasoning-Driven Generation, Retrieval and Compression 2026-05-12 conf:96% GRC unifies generation, retrieval-oriented representation, and context compression in one LLM forward pass. It introduces latent memory-augmented generation with compressed updatable KV-cache memory, reducing RAG deployment cost while improving efficiency on retrieval, generation, and compression tasks.

• FibQuant: Universal Vector Quantization for Random-Access KV-Cache Compression 2026-05-12 conf:94% FibQuant compresses LLM KV caches with a universal random-access vector quantizer that better preserves attention quality than prior scalar methods. It enables much higher context-memory efficiency, achieving strong compression with minimal perplexity loss on GPT-2 and TinyLlama.

• Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression 2026-05-12 conf:91% The paper introduces KVFundaBench to evaluate how KV cache compression affects long-context reasoning, showing that reasoning degrades more than retrieval under aggressive compression. It proposes ShotKV, which preserves semantic units like few-shot examples during cache compression, improving accuracy and reducing latency.

• Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live 2026-05-11 conf:94% Continuum introduces CacheTTL, a KV-cache time-to-live mechanism that keeps agent conversation state in GPU memory across tool calls and evicts it when reuse is unlikely. This improves multi-turn agent latency and throughput by making context reuse more efficient and robust.

• Not All Thoughts Need HBM: Semantics-Aware Memory Hierarchy for LLM Reasoning 2026-05-10 conf:98% This paper proposes a semantics-aware memory hierarchy for LLM reasoning that offloads low-importance KV-cache tokens from GPU HBM to CPU memory instead of evicting them. It shows that preserving tokens via zero-error offloading keeps reasoning accuracy high while substantially reducing GPU memory use with modest transfer overhead.

• KV-RM: Regularizing KV-Cache Movement for Static-Graph LLM Serving 2026-05-10 conf:82% KV-RM improves static-graph LLM serving by regularizing KV-cache movement with paged storage, committed decode descriptors, and coalesced transfers. This reduces reserved memory, improves mixed-length decoding throughput, and cuts tail-latency spikes.

• ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing 2026-05-09 conf:98% ReST-KV improves long-context LLM inference by evicting KV-cache entries using layer-wise output reconstruction plus spatial-temporal smoothing. It models how token removal changes outputs, yielding better accuracy on long-context benchmarks and much lower decoding latency.

• Reformulating KV Cache Eviction Problem for Long-Context LLM Inference 2026-05-08 conf:98% This paper reframes KV cache eviction for long-context LLMs as an output-aware approximation problem and proposes LaProx, a global token scoring method that better estimates cache importance. It preserves performance with only 5% of the KV cache and outperforms prior eviction methods on long-context benchmarks.

• WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki Systems 2026-05-08 conf:98% WiCER improves persistent wiki-style memory for LLMs by iteratively detecting facts lost during knowledge compilation and forcing them to be preserved. It shows that this refinement process reduces catastrophic failures and recovers most of the quality lost when compressing raw documents into fast KV-cache-served memory.

• RDKV: Rate-Distortion Bit Allocation for Joint Eviction and Quantization of the KV Cache 2026-05-08 conf:95% RDKV treats KV-cache compression as a rate-distortion problem, jointly deciding token/channel eviction and quantization instead of handling them separately. It greatly reduces long-context memory and decoding cost while preserving near full-cache accuracy.

• CSR: Infinite-Horizon Real-Time Policies with Massive Cached State Representations 2026-05-08 conf:94% The paper proposes Cached State Representation and Asynchronous State Reconciliation to let LLMs maintain massive long-horizon state with efficient KV-cache reuse and bounded latency. It shows large latency reductions and strong recall for real-time embodied policies.

• Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models 2026-05-08 conf:88% MELT introduces a looped transformer architecture that keeps a single shared KV cache across reasoning iterations, using gating to update it and achieve constant-memory iterative reasoning. It matches or exceeds comparable models while dramatically reducing memory use compared with prior looped LLMs.

• SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving 2026-05-08 conf:88% SplitZip is a GPU-friendly lossless compressor for KV-cache transfer in disaggregated LLM serving. It speeds up moving short-term attention memory between prefill and decode workers, reducing latency and improving throughput for long-context and agentic workloads.

• Priming: Hybrid State Space Models From Pre-trained Transformers 2026-05-08 conf:81% Priming converts pre-trained Transformers into hybrid attention-SSM models with stronger long-context performance, smaller KV caches, and faster decoding using minimal extra training. It also provides a scalable comparison of SSM memory layers and releases open-source models and tooling.

• When Does Value-Aware KV Eviction Help? A Fixed-Contract Diagnostic for Non-Monotone Cache Compression 2026-05-07 conf:93% This paper analyzes when value-aware KV-cache eviction improves long-context LLM inference. It proposes a diagnostic that isolates cache-selection errors and shows effective compression should first preserve needed evidence, then rank its output importance, and avoid breaking coupled evidence.

• How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment 2026-05-07 conf:92% This paper studies how KV-cache compression affects long-context RL rollouts for LLMs, where small cache errors can create large off-policy bias. It targets memory-efficient alignment by addressing the tradeoff between reduced memory use and stable RL optimization.

• HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding 2026-05-07 conf:89% HERMES proposes treating the KV cache of a multimodal LLM as a hierarchical memory for streaming video understanding. It reuses a compact cache to preserve relevant past video information, enabling faster real-time responses and lower memory usage while maintaining or improving accuracy on streaming benchmarks.

• When Quantization Is Free: An int4 KV Cache That Outruns fp16 on Apple Silicon 2026-05-07 conf:87% The paper introduces an int4 KV-cache system for LLMs on Apple Silicon that is faster than fp16 while cutting persistent cache memory by 3x. It uses a fused Metal kernel and shows near-preserved quality, making longer-context inference more efficient.

• DUAL-BLADE: Dual-Path NVMe-Direct KV-Cache Offloading for Edge LLM Inference 2026-04-29 conf:82% DUAL-BLADE improves edge LLM inference by offloading KV cache to NVMe with a dual-path design that ch