Memory Systems Survey

Introduction

Memory is the cornerstone of intelligence. Without memory, agents cannot learn from experience, maintain conversational coherence, or behave consistently in complex tasks. This section starts from psychological models of human memory to establish a conceptual framework for agent memory systems.

Human Memory Models

The Atkinson-Shiffrin Model (Multi-Store Model)

In 1968, Atkinson and Shiffrin proposed the classic multi-store model, dividing human memory into three stages:

1. Sensory Memory: Extremely short duration (<1 second), large capacity but rapid decay
2. Short-Term Memory: Limited capacity (7 plus or minus 2 items), lasts about 20-30 seconds
3. Long-Term Memory: Nearly unlimited capacity, can persist for years or even a lifetime

graph LR
    A[External Stimuli] --> B[Sensory Memory<br/>< 1 second]
    B -->|Attention| C[Short-Term Memory<br/>7±2 items<br/>20-30 seconds]
    B -->|Unattended| D[Forgetting]
    C -->|Encoding/Rehearsal| E[Long-Term Memory<br/>Unlimited capacity<br/>Persistent storage]
    C -->|Unrehearsed| D
    E -->|Retrieval| C

Baddeley's Working Memory Model

Baddeley (1974, 2000) extended the concept of short-term memory, proposing that working memory contains:

• Central Executive: Attention control and coordination
• Phonological Loop: Temporary storage of verbal information
• Visuospatial Sketchpad: Visual and spatial information processing
• Episodic Buffer: Integration of multimodal information

Tulving's Long-Term Memory Classification

• Episodic Memory: Personal experiences and events
• Semantic Memory: General knowledge and facts
• Procedural Memory: Skills and operational procedures

From Human Memory to Agent Memory

Mapping Relationships

Human Memory	Agent Memory	Implementation
Sensory memory	Input buffer	Raw input (user messages, API returns)
Short-term/Working memory	Context window	LLM's context window (4K-128K tokens)
Long-term - Episodic	Conversation history/Experience store	Historical interactions stored in vector databases
Long-term - Semantic	Knowledge base	Knowledge graphs, structured databases
Long-term - Procedural	Learned workflows	Stored prompt templates, tool call patterns

Core Challenges of Agent Memory

1. Capacity limitations: Context window is finite; cannot remember everything
2. Retrieval efficiency: How to quickly find relevant memories
3. Memory updates: How to handle outdated or contradictory information
4. Importance assessment: Which information is worth long-term storage

Agent Memory Architecture

graph TB
    subgraph Input Layer
        U[User Input] --> WM
        T[Tool Returns] --> WM
        E[Environment Observations] --> WM
    end

    subgraph Working Memory
        WM[Context Window]
    end

    subgraph Long-Term Memory
        WM -->|Store| VDB[(Vector Database<br/>Episodic Memory)]
        WM -->|Store| KG[(Knowledge Graph<br/>Semantic Memory)]
        WM -->|Store| PM[(Workflow Library<br/>Procedural Memory)]
    end

    subgraph Retrieval
        VDB -->|Similarity Search| R[Retrieval Module]
        KG -->|Structured Query| R
        PM -->|Pattern Matching| R
        R -->|Inject into Context| WM
    end

    WM --> LLM[LLM Reasoning]
    LLM --> OUT[Output/Action]

Memory Type Details

By Time Span

Type	Time Span	Manifestation in Agents
Immediate memory	Current turn	Information in current prompt
Short-term memory	Current session	Conversation history (within session)
Medium-term memory	Cross-session	User preferences, task progress
Long-term memory	Permanent	Knowledge base, learned skills

By Content Type

• Factual memory: Specific information the user tells the agent ("My name is...")
• Procedural memory: Operational patterns the agent has learned ("For this type of problem, first...")
• Meta-memory: Memory about memory itself ("This information was discussed before")

Evaluation Metrics for Memory Systems

• Recall: Can relevant memories be retrieved?
• Precision: Are retrieved memories relevant?
• Latency: Retrieval speed
• Consistency: Are memories free of contradictions?
• Freshness: Does it reflect the most current information?

Chapter Structure

This chapter explores every aspect of agent memory systems in depth:

1. Working Memory and Context Management - Efficient utilization of the context window
2. Long-Term Memory and Vector Databases - Persistent storage and retrieval
3. RAG-Augmented Memory - Complete retrieval-augmented generation pipeline
4. Episodic and Semantic Memory - Design of different memory types
5. Memory Architecture Design - Comprehensive architecture and frontier approaches

References

• Atkinson, R. C., & Shiffrin, R. M. (1968). "Human memory: A proposed system and its control processes"
• Baddeley, A. (2000). "The episodic buffer: A new component of working memory?"
• Tulving, E. (1972). "Episodic and semantic memory"
• Sumers, T. R., et al. (2023). "Cognitive Architectures for Language Agents"
• Zhang, Z., et al. (2024). "A Survey on the Memory Mechanism of Large Language Model based Agents"

---

评论