FeishuLuo/Evolving-LLM-Agent-Memory-Survey

Paper list for "From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms".

56

48 commits

updated Apr 13, 2026

See the code

README

Evolution Path of LLM Agent Memory

From Storage to Experience:
A Survey on the Evolution of LLM Agent Memory Mechanisms

Paper GitHub Stars License Last Commit Issues Paper Count PRs Welcome

ACL 2026 (Findings)  |  ICLR 2026 Workshop MemAgents  |  Actively maintained

A curated and continuously updated collection of 140+ papers and 40+ benchmarks
on the evolutionary framework of LLM agent memory mechanisms.

If you find this survey useful, please consider giving us a :star: to stay updated with the latest additions!

We welcome contributions! If you know of a relevant paper we missed, please open an issue or submit a pull request. See our Contributing Guidelines.


Table of Contents


Overview

While memory mechanisms have emerged as the architectural cornerstone of LLM agents, current research remains fragmented between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis.

We propose a novel Evolutionary Framework that formalizes the development of LLM agent memory into three progressive stages:

StageCore IdeaKey Transformation
StorageTrajectory PreservationFaithfully recording raw interaction traces via linear, vector, or structured storage
ReflectionTrajectory RefinementActively evaluating and correcting stored memories through introspection, environment feedback, or coordination
ExperienceTrajectory AbstractionCompressing redundant trajectories into transferable heuristic wisdom via cross-trajectory abstraction

Evolution Path

The agent's decision-making process forms a dynamic closed loop enabled by memory. During each task execution cycle, the agent leverages two core capabilities:

  1. Memory Read — Actively retrieves relevant knowledge from the memory bank to supplement the current context.
  2. Memory Write — Records generated interaction sequences as historical trajectories into the memory system.

Building upon this foundation, we conceptualize the memory mechanism as an evolutionary pathway structured into three stages:


Storage (Trajectory Preservation)

Storage serves as the cornerstone of memory evolution, emphasizing the faithful preservation of interaction history.

ParadigmCharacteristicsStrengthsLimitations
LinearToken stream ordered chronologicallyMinimal information loss, maximal logical completenessEarly critical information irreversibly forgotten
VectorEmbeddings in high-dimensional spaceMassive storage capacityHigh retrieval difficulty, limited relevance
StructuredRelational structures (tables, graphs, tiers)Precise operations, multi-hop retrievalRequires schema maintenance, less flexible at scale
Reflection (Trajectory Refinement)

Drivers in Dynamic Environments
Figure 2: Temporal Validity & Causal Structure in Dynamic Environments.

ParadigmCharacteristicsStrengthsLimitations
IntrospectionInternal knowledge for self-evaluationError correction without external feedbackRisk of reinforcing biases
EnvironmentExecution outcomes as refinement signalsGreater adaptability to dynamic environmentsSparse rewards, ambiguous settings
CoordinationMulti-agent collective reflectionReduced hallucination, enriched perspectivesCommunication overhead, memory conflicts
Experience (Trajectory Abstraction)

Cross-Trajectory Abstraction
Figure 3: Overview of Cross-Trajectory Abstraction Techniques.

ParadigmCharacteristicsStrengthsLimitations
ExplicitHuman-readable patterns from trajectory clustersHighly interpretable and editableLacks precision in complex boundaries
ImplicitInternalized into model parameters / latent variablesNear-zero retrieval overheadReduced interpretability, forgetting risk
HybridDynamic "accumulate–internalize" cycleBalances interpretability with efficiencyRequires careful transfer coordination

Paper List

Papers are organized following our three-stage evolutionary framework.

Storage

Linear Storage

Context Window Adaptation

Information Sparsification

Vector Storage

Semantic Retrieval

Weighted Retrieval

Structured Storage

Tabular Database

Tiered Architectures

Semantic Graphs


Reflection

Introspection

Error Rectification

Dynamic Maintenance

Knowledge Compression

Environment

Environment Modeling

Decision Optimization

Coordination

Multi-dimensional Calibration


Experience

Explicit Experience

Heuristic Guidelines

Procedural Primitives

Implicit Experience

Latent Modulation

Parameter Internalization

Hybrid Experience

Experience Transfer


Benchmarks & Datasets

Storage Stage

BenchmarkVenueYearFocus
HotpotQAEMNLP2018Multi-hop cross-document reasoning
LongBencharXiv2023Bilingual, multitask long-context understanding
MemoryBankarXiv2023Long-term dialog memory
Multimodal Needle in a HaystackNAACL2024Long-context multimodal retrieval
LongBench v2arXiv2024Realistic long-context reasoning
RULERarXiv2024Effective context window measurement
BABILongarXiv2024Synthetic long-context reasoning-in-a-haystack
DialSimarXiv2024Real-time long-term dialogue simulation
Evaluating Very Long-Term Conversational Memory of LLM AgentsarXiv2024Very long-term conversational memory
MADial-BenchNAACL2024Memory-augmented dialogue
HELMETarXiv2024Comprehensive long-context evaluation
Explicit vs. Implicit MemoryarXiv2025Multi-hop personalized reasoning
Evaluating the Long-Term Memory of Large Language ModelsACL2025Long-term memory evaluation protocol
Beyond a Million TokensarXiv2025Ultra-long memory (>1M tokens)
LoCoBench-AgentarXiv2025Long-context software engineering agents
MemoryRewardBencharXiv2026Reward models for long-term memory management
AgentLongBencharXiv2026Dynamic long-context agent evaluation

Reflection Stage

BenchmarkVenueYearFocus
Personalized Large Language Model Assistant with Evolving Conditional MemoryCOLING2023Evolving conditional memory
PerLTQAarXiv2024Personal long-term memory QA
Evaluating Very Long-Term Conversational Memory of LLM AgentsarXiv2024Conversational memory consistency
On the Multi-Turn Instruction Following for Conversational Web AgentsACL2024Web agent multi-turn memory
SHAREarXiv2024Shared memory in dialogue
MinervaarXiv2025Programmable memory read-write tests
Personalized Preference FollowingarXiv2025User preference memory
Multi-Session Personalized ConversationarXiv2025Multi-session implicit reasoning
PersonaMem-v2arXiv2025Implicit user persona memory
Mem-PALarXiv2025Personalized long-term dialogue
WebChoreArenaarXiv2025Web agent intermediate state memory
StoryBencharXiv2025Long-term narrative tracking
Evaluating Memory in LLM Agents via Incremental Multi-Turn InteractionsarXiv2025Memory retention across turns
LLM Self-Awareness via Internal CircuitsarXiv2025Internal memory limit awareness
HaluMemarXiv2025Hallucination in memory systems
ConvoMem BenchmarkarXiv2025Conversational memory (75K+ QA pairs)
KnowMe-BencharXiv2026Person understanding from narratives
RealMemarXiv2026Real-world project-oriented memory
StructMemEvalarXiv2026Memory structure organization
Beyond the Context Window: Cost-Performance Analysis of Fact-Based Memory vs. Long-Context LLMsarXiv2026Fact-based memory vs. long-context cost tradeoff
Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent MemoryarXiv2026Write strategy x retrieval method diagnostics

Experience Stage

BenchmarkVenueYearFocus
StreamBencharXiv2024Continuous improvement over task streams
LifelongAgentBencharXiv2025Lifelong learning agents
MEMTRACKNeurIPS Workshop2025Multi-platform state tracking
MemoryBencharXiv2025Memory and continual learning metrics
Evo-MemoryarXiv2025Test-time self-evolving memory
MemoryArenaarXiv2026Interdependent multi-session agentic tasks
AMA-BencharXiv2026Long-horizon agentic memory

Citation

If you find this survey useful in your research, please consider citing our paper:

@inproceedings{luo2026from,
  title   = {From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms},
  author  = {Luo, Jinghao and Tian, Yuchen and Cao, Chuxue and Luo, Ziyang and Lin, Hongzhan and Li, Kaixin and Kong, Chuyi and Yang, Ruichao and Ma, Jing},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year    = {2026}
}

The BibTeX entry will be updated with the official ACL 2026 proceedings metadata once available.


Contributing

We welcome contributions from the community! Please see our Contributing Guidelines for how to:

  • Suggest new papers
  • Report broken links
  • Propose new categories

This project is licensed under the MIT License.

Contributors

FeishuLuo

46 commits

ChiYeungLaw

2 commits

FeishuLuo/Evolving-LLM-Agent-Memory-Survey

Paper list for "From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms".

56

48 commits

updated Apr 13, 2026

See the code

README

Evolution Path of LLM Agent Memory

From Storage to Experience:
A Survey on the Evolution of LLM Agent Memory Mechanisms

Paper GitHub Stars License Last Commit Issues Paper Count PRs Welcome

ACL 2026 (Findings)  |  ICLR 2026 Workshop MemAgents  |  Actively maintained

A curated and continuously updated collection of 140+ papers and 40+ benchmarks
on the evolutionary framework of LLM agent memory mechanisms.

If you find this survey useful, please consider giving us a :star: to stay updated with the latest additions!

We welcome contributions! If you know of a relevant paper we missed, please open an issue or submit a pull request. See our Contributing Guidelines.


Table of Contents


Overview

While memory mechanisms have emerged as the architectural cornerstone of LLM agents, current research remains fragmented between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis.

We propose a novel Evolutionary Framework that formalizes the development of LLM agent memory into three progressive stages:

StageCore IdeaKey Transformation
StorageTrajectory PreservationFaithfully recording raw interaction traces via linear, vector, or structured storage
ReflectionTrajectory RefinementActively evaluating and correcting stored memories through introspection, environment feedback, or coordination
ExperienceTrajectory AbstractionCompressing redundant trajectories into transferable heuristic wisdom via cross-trajectory abstraction

Evolution Path

The agent's decision-making process forms a dynamic closed loop enabled by memory. During each task execution cycle, the agent leverages two core capabilities:

  1. Memory Read — Actively retrieves relevant knowledge from the memory bank to supplement the current context.
  2. Memory Write — Records generated interaction sequences as historical trajectories into the memory system.

Building upon this foundation, we conceptualize the memory mechanism as an evolutionary pathway structured into three stages:


Storage (Trajectory Preservation)

Storage serves as the cornerstone of memory evolution, emphasizing the faithful preservation of interaction history.

ParadigmCharacteristicsStrengthsLimitations
LinearToken stream ordered chronologicallyMinimal information loss, maximal logical completenessEarly critical information irreversibly forgotten
VectorEmbeddings in high-dimensional spaceMassive storage capacityHigh retrieval difficulty, limited relevance
StructuredRelational structures (tables, graphs, tiers)Precise operations, multi-hop retrievalRequires schema maintenance, less flexible at scale
Reflection (Trajectory Refinement)

Drivers in Dynamic Environments
Figure 2: Temporal Validity & Causal Structure in Dynamic Environments.

ParadigmCharacteristicsStrengthsLimitations
IntrospectionInternal knowledge for self-evaluationError correction without external feedbackRisk of reinforcing biases
EnvironmentExecution outcomes as refinement signalsGreater adaptability to dynamic environmentsSparse rewards, ambiguous settings
CoordinationMulti-agent collective reflectionReduced hallucination, enriched perspectivesCommunication overhead, memory conflicts
Experience (Trajectory Abstraction)

Cross-Trajectory Abstraction
Figure 3: Overview of Cross-Trajectory Abstraction Techniques.

ParadigmCharacteristicsStrengthsLimitations
ExplicitHuman-readable patterns from trajectory clustersHighly interpretable and editableLacks precision in complex boundaries
ImplicitInternalized into model parameters / latent variablesNear-zero retrieval overheadReduced interpretability, forgetting risk
HybridDynamic "accumulate–internalize" cycleBalances interpretability with efficiencyRequires careful transfer coordination

Paper List

Papers are organized following our three-stage evolutionary framework.

Storage

Linear Storage

Context Window Adaptation

Information Sparsification

Vector Storage

Semantic Retrieval

Weighted Retrieval

Structured Storage

Tabular Database

Tiered Architectures

Semantic Graphs


Reflection

Introspection

Error Rectification

Dynamic Maintenance

Knowledge Compression

Environment

Environment Modeling

Decision Optimization

Coordination

Multi-dimensional Calibration


Experience

Explicit Experience

Heuristic Guidelines

Procedural Primitives

Implicit Experience

Latent Modulation

Parameter Internalization

Hybrid Experience

Experience Transfer


Benchmarks & Datasets

Storage Stage

BenchmarkVenueYearFocus
HotpotQAEMNLP2018Multi-hop cross-document reasoning
LongBencharXiv2023Bilingual, multitask long-context understanding
MemoryBankarXiv2023Long-term dialog memory
Multimodal Needle in a HaystackNAACL2024Long-context multimodal retrieval
LongBench v2arXiv2024Realistic long-context reasoning
RULERarXiv2024Effective context window measurement
BABILongarXiv2024Synthetic long-context reasoning-in-a-haystack
DialSimarXiv2024Real-time long-term dialogue simulation
Evaluating Very Long-Term Conversational Memory of LLM AgentsarXiv2024Very long-term conversational memory
MADial-BenchNAACL2024Memory-augmented dialogue
HELMETarXiv2024Comprehensive long-context evaluation
Explicit vs. Implicit MemoryarXiv2025Multi-hop personalized reasoning
Evaluating the Long-Term Memory of Large Language ModelsACL2025Long-term memory evaluation protocol
Beyond a Million TokensarXiv2025Ultra-long memory (>1M tokens)
LoCoBench-AgentarXiv2025Long-context software engineering agents
MemoryRewardBencharXiv2026Reward models for long-term memory management
AgentLongBencharXiv2026Dynamic long-context agent evaluation

Reflection Stage

BenchmarkVenueYearFocus
Personalized Large Language Model Assistant with Evolving Conditional MemoryCOLING2023Evolving conditional memory
PerLTQAarXiv2024Personal long-term memory QA
Evaluating Very Long-Term Conversational Memory of LLM AgentsarXiv2024Conversational memory consistency
On the Multi-Turn Instruction Following for Conversational Web AgentsACL2024Web agent multi-turn memory
SHAREarXiv2024Shared memory in dialogue
MinervaarXiv2025Programmable memory read-write tests
Personalized Preference FollowingarXiv2025User preference memory
Multi-Session Personalized ConversationarXiv2025Multi-session implicit reasoning
PersonaMem-v2arXiv2025Implicit user persona memory
Mem-PALarXiv2025Personalized long-term dialogue
WebChoreArenaarXiv2025Web agent intermediate state memory
StoryBencharXiv2025Long-term narrative tracking
Evaluating Memory in LLM Agents via Incremental Multi-Turn InteractionsarXiv2025Memory retention across turns
LLM Self-Awareness via Internal CircuitsarXiv2025Internal memory limit awareness
HaluMemarXiv2025Hallucination in memory systems
ConvoMem BenchmarkarXiv2025Conversational memory (75K+ QA pairs)
KnowMe-BencharXiv2026Person understanding from narratives
RealMemarXiv2026Real-world project-oriented memory
StructMemEvalarXiv2026Memory structure organization
Beyond the Context Window: Cost-Performance Analysis of Fact-Based Memory vs. Long-Context LLMsarXiv2026Fact-based memory vs. long-context cost tradeoff
Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent MemoryarXiv2026Write strategy x retrieval method diagnostics

Experience Stage

BenchmarkVenueYearFocus
StreamBencharXiv2024Continuous improvement over task streams
LifelongAgentBencharXiv2025Lifelong learning agents
MEMTRACKNeurIPS Workshop2025Multi-platform state tracking
MemoryBencharXiv2025Memory and continual learning metrics
Evo-MemoryarXiv2025Test-time self-evolving memory
MemoryArenaarXiv2026Interdependent multi-session agentic tasks
AMA-BencharXiv2026Long-horizon agentic memory

Citation

If you find this survey useful in your research, please consider citing our paper:

@inproceedings{luo2026from,
  title   = {From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms},
  author  = {Luo, Jinghao and Tian, Yuchen and Cao, Chuxue and Luo, Ziyang and Lin, Hongzhan and Li, Kaixin and Kong, Chuyi and Yang, Ruichao and Ma, Jing},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year    = {2026}
}

The BibTeX entry will be updated with the official ACL 2026 proceedings metadata once available.


Contributing

We welcome contributions from the community! Please see our Contributing Guidelines for how to:

  • Suggest new papers
  • Report broken links
  • Propose new categories

This project is licensed under the MIT License.

Contributors

FeishuLuo

46 commits

ChiYeungLaw

2 commits