🧠 Memory is Awesome

Your Guide to Memory in Foundation Model Agents

"Without memory, there is no culture. Without memory, there would be no civilization, no society, no future."
— Elie Wiesel

---

📖 Table of Contents

• Introduction
• What is Agent Memory?
• Unified Taxonomy
  • Memory Forms
  • Memory Functions
  • Memory Dynamics
• Paper Collection
  • Factual Memory
  • Experiential Memory
  • Working Memory
• Benchmarks & Evaluation
• Open-Source Frameworks
• Applications
• Future Directions
• References
• Citation
• Contributing

---

🎯 Introduction

Foundation model-based agents have emerged as a transformative paradigm in AI research. Unlike vanilla foundation models, these agents possess self-evolving capabilities that enable them to solve real-world problems requiring long-term, complex interactions with their environment.

Memory is the cornerstone of these agents—it's what makes an agent truly an agent. Memory underpins the ability to perform long-horizon reasoning, adapt continually, and interact effectively with complex environments.

📊 Repository Highlights

This repository collects agent memory research, featuring:

• 350+ papers spanning from foundational works (Neural Turing Machines, 2014) to recent research (March 2026)
• Unified taxonomy organizing research by Forms × Functions × Dynamics
• 11+ benchmarks for evaluating memory capabilities (LoCoMo, LongMemEval, MemBench, LaMP, etc.)
• 9+ open-source frameworks (Mem0, A-MEM, Zep, MemGPT, HippoRAG, etc.)
• 13+ references including surveys synthesizing the field's evolution

🔬 Coverage Areas

This repository synthesizes insights from surveys on agent memory (see References).

---

🧩 What is Agent Memory?

Conceptual Distinction

Agent Memory is distinct from related concepts:

What Counts as Memory?

---

🗂️ Unified Taxonomy

This repository organizes agent memory research through three unified lenses: Forms, Functions, and Dynamics.

---

📦 Memory Forms (What Carries Memory?)

Memory forms categorize HOW information is stored and represented.

🔤 Token-level Memory

💡 Why use it? Token-level memory is human-readable and interpretable, making it easy to debug, update, and audit. It allows for flexible retrieval strategies and can be shared across different models without retraining.

Subtypes:

⚙️ Parametric Memory

💡 Why use it? Parametric memory enables permanent knowledge storage that doesn't require retrieval at inference time. It's ideal for domain-specific adaptation and can capture nuanced patterns that are difficult to express in text.

🧬 Latent Memory

💡 Why use it? Latent memory provides highly efficient storage with minimal overhead. It's particularly effective for working memory scenarios where information needs to be quickly accessed and doesn't need to be human-interpretable.

---

🎯 Memory Functions (Why Agents Need Memory?)

Memory functions categorize WHAT purpose the memory serves.

---

🔄 Memory Dynamics (How Memory Evolves?)

Memory dynamics describe the operational lifecycle of memory.

---

📚 Paper Collection

📚 Factual Memory

Factual Memory (also called Semantic Memory) stores general world knowledge, facts, and information about entities and their relationships. It answers "what do I know?" and provides the knowledge base that agents draw upon for reasoning and question-answering.

🔤 Token-level Factual Memory

Token-level Factual Memory stores world knowledge, facts, and semantic information as explicit text or structured data (e.g., knowledge graphs, databases, retrieved documents). This is the most common form of external memory in RAG systems, where facts are retrieved as text chunks and injected into the context window.

⚙️ Parametric Factual Memory

Parametric Factual Memory encodes factual knowledge directly into model weights through training or fine-tuning. This includes knowledge editing methods (ROME, MEMIT), adapter-based knowledge injection (K-Adapter, LoRA), and continual learning approaches that update model parameters to incorporate new facts.

🧬 Latent Factual Memory

Latent Factual Memory stores factual information in continuous vector representations or hidden states. This includes memory-augmented architectures (Neural Turing Machines, Memorizing Transformers), state space models (Mamba, RWKV), and learned memory tokens that compress knowledge into dense representations.

🎓 Experiential Memory

Experiential Memory (also called Episodic or Procedural Memory) stores records of past experiences, interactions, and learned skills. It answers "what have I done?" and enables agents to learn from past successes and failures, accumulate skills, and improve over time.

🔤 Token-level Experiential Memory

Token-level Experiential Memory stores past experiences, interactions, and learned skills as explicit records (e.g., conversation logs, action trajectories, skill libraries). Agents retrieve relevant past experiences to inform current decisions, enabling learning from trial-and-error and skill accumulation.

⚙️ Parametric Experiential Memory

Parametric Experiential Memory encodes learned experiences and skills into model parameters through reinforcement learning, imitation learning, or continual fine-tuning. This includes RLHF, policy gradient methods, and approaches that update model weights based on interaction feedback.

🧬 Latent Experiential Memory

Latent Experiential Memory stores experiences as latent representations, such as experience replay buffers in RL or learned skill embeddings. These compressed representations enable efficient storage and generalization across similar experiences.

⚡ Working Memory

Working Memory (also called Short-term Memory) manages the currently active context and information being processed. It answers "what am I focusing on now?" and handles the limited attention window, deciding what to keep, compress, or discard during extended interactions.

🔤 Token-level Working Memory

Token-level Working Memory manages the active context window through explicit text manipulation—deciding what information to keep, summarize, or discard as conversations extend beyond context limits. This includes context compression, summarization, and selective attention mechanisms.

⚙️ Parametric Working Memory

Parametric Working Memory implements working memory through learned model components, such as attention mechanisms that learn what to focus on, or architectural modifications that improve context utilization efficiency.

🧬 Latent Working Memory

Latent Working Memory manages active context through compressed latent representations, including KV cache optimization, recurrent memory states, and memory tokens. These approaches reduce memory footprint while preserving essential information for ongoing computation.

📊 Benchmarks & Evaluation

Memory Evaluation Benchmarks

Evaluation Metrics

---

🛠️ Open-Source Frameworks

---

🎮 Applications

Memory mechanisms are used across various LLM agent applications:

---

🔮 Future Directions

Based on current research, future directions include:

---

📚 References

This repository synthesizes insights from the following surveys and papers:

Surveys

• Memory in the Age of AI Agents: A Survey (2025) [GitHub]
• Memory-Augmented Transformers: From Neuroscience Principles to Technical Solutions (2025)
• From S4 to Mamba: A Comprehensive Survey on Structured State Space Models (2025)
• From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs (2025)
• Retrieval-Augmented Generation for Natural Language Processing: A Survey (2025)
• A Survey of Robotic Navigation and Manipulation with Physics Simulators in the Era of Embodied AI (2025)
• KV Cache Compression for Inference Efficiency in LLMs: A Survey (2024)
• A Comprehensive Survey of Continual Learning: Theory, Method and Application (2024) [GitHub]
• Retrieval-Augmented Generation for Large Language Models: A Survey (2024)
• A Survey on the Memory Mechanism of Large Language Model based Agents (2024) [GitHub]
• Robot Learning in the Era of Foundation Models: A Survey (2023)
• Memory in LLM-based Multi-agent Systems: Mechanisms, Challenges, and Collective Intelligence (2025)
• Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers (2026)
• Continual Learning in LLMs: Methods, Challenges, and Opportunities (2026)

Papers

• Continual Learning as Computationally Constrained Reinforcement Learning (2023)

---

📄 Citation

If you find this repository helpful, please cite:

@misc{memoryisawesome,
  title={Memory is Awesome: Your Guide to Memory in Foundation Model Agents},
  author={SuperMadee},
  year={2024},
  howpublished={\url{https://github.com/SuperMadee/MemoryIsAwesome}}
}

---

🤝 Contributing

Contributions are welcome! If you'd like to add new papers, fix errors, or suggest improvements:

1. Fork the repository
2. Create a new branch (git checkout -b feature/add-paper)
3. Make your changes
4. Submit a pull request

Please ensure any added papers include:

• Full paper title with year
• Links to arXiv paper AND GitHub (if available)
• Appropriate categorization by Form × Function

---

⭐ Star this repo if you find it helpful!

Made with ❤️ for the Agent Research Community