A Survey on the Memory Mechanism of Large Language Model-based Agents

📝 Paper Summary

Memory recall Memory organization Self-evolving Agentic reasoning

This survey provides a comprehensive taxonomy and analysis of memory mechanisms in LLM-based agents, categorizing them by information sources, storage forms, and operations to guide future design.

Core Problem

While many LLM-based agents implement memory to handle long-term interactions, existing research is scattered across isolated papers without a unified framework or design patterns.

Why it matters:

Without memory, agents cannot solve real-world problems requiring long-term context, such as booking tickets that depend on prior feedback or maintaining personal assistant continuity
Current ad-hoc memory designs lack standardization, making it difficult for researchers to compare mechanisms or abstract common effective patterns
The transition from static LLMs to self-evolving AGI requires agents to autonomously accumulate knowledge and learn from experience, which is impossible without structured memory

Concrete Example: A trip-planning agent without memory sends a ticket request but forgets the response before taking the next action, resulting in a failure to book. A memory-enhanced agent stores the website's response, retrieves it, and proceeds to the payment step.

Key Novelty

Unified Taxonomy for Agent Memory

Deconstructs agent memory into three core dimensions: Sources (where info comes from), Forms (how it is stored), and Operations (how it is manipulated)
Distinguishes between 'Short-term' (context window) and 'Long-term' (external storage/parameters) memory in the context of cognitive psychology analogies
Categorizes evaluation methods into Direct Evaluation (measuring memory retention) and Indirect Evaluation (measuring downstream task performance)

Evaluation Highlights

Reviews memory mechanisms across diverse applications including role-playing, social simulation, personal assistants, and code generation
Identifies that memory enhances self-evolution by allowing agents to reflect on past experiences (e.g., Reflexion) and adjust future behaviors
Highlights the trade-off between memory retrieval accuracy and the computational cost of maintaining large external memory banks

Breakthrough Assessment

7/10

A timely and necessary systematization of a rapidly growing field. While it is a survey (no new algorithm), its taxonomy provides a crucial foundation for future agent memory research.

⚙️ Technical Details

Problem Definition

Setting: Designing memory modules for LLM-based agents to enable experience accumulation and self-evolution

Inputs: Observations from the environment, user feedback, and agent's internal thoughts

Outputs: Stored memory records, retrieved contexts, and updated agent behaviors

Pipeline Flow

Memory Acquisition (Input from Env/User/Self)
Memory Storage (Writing to Database/Buffer)
Memory Retrieval (Reading for Context)
Memory Evolution (Reflection/Forgetting)

System Modules

Memory Reader (Memory Operations)

Retrieves relevant information from storage to support current decision making

Model or implementation: Retriever (e.g., Vector Search, Keyword Match)

Memory Writer (Memory Operations)

Stores new experiences into the memory system

Model or implementation: Database Writer / LLM Summarizer

Memory Storage

Holds the accumulated knowledge

Model or implementation: Vector Database / Key-Value Store / Text Log

Novel Architectural Elements

Taxonomy separating memory into Sources (User, Environment, Agent), Forms (Text, Vector, Database, Parametric), and Operations (Read, Write, Reflect)
Unified view of 'Memory' as the bridge between static LLM capabilities and dynamic Agentic behaviors

Comparison to Prior Work

vs. Reflexion: This paper surveys Reflexion as a specific instance of memory operation (reflection) rather than proposing it
vs. Generative Agents: This paper categorizes the architecture of Generative Agents into its broader taxonomy (Textual Form, User/Agent Source, Retrieval Operation)
Novel contribution: Provides the first holistic survey specifically focused on the *memory mechanism* of agents, whereas previous surveys focused on agents broadly [survey_llm_agent1]

Limitations

Existing memory retrieval is often inaccurate or biased by recency
Memory storage capacity is limited by retrieval latency and cost
Lack of standardized evaluation benchmarks for memory modules specifically
Privacy concerns regarding storing user interactions in long-term memory

Reproducibility

Code: https://github.com/nuster1128/LLM_Agent_Memory_Survey

The paper is a survey; the 'code availability' refers to a GitHub repository collecting the surveyed papers. No specific model training code is applicable to this survey itself.

📊 Experiments & Results

Evaluation Setup

Review of evaluation methodologies used in the surveyed papers

Benchmarks:

Direct Evaluation (Memory Recall/Retention)
Indirect Evaluation (Downstream Agent Tasks (e.g., Game Playing, Coding))

Metrics:

Retrieval Accuracy
Task Success Rate
User Satisfaction
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Memory is the key component enabling 'self-evolution' in agents, allowing them to improve over time without retraining weights
Most current agents use textual or vector-based external memory; parametric memory (updating weights) is less common due to cost and catastrophic forgetting risks
Evaluation is bifurcated: some papers test memory directly (can you recall X?), but most rely on indirect task performance (did the agent win the game because it remembered X?)
Future directions include more efficient memory structures (hierarchical/graph-based) and better forgetting mechanisms to manage information overload

📚 Prerequisite Knowledge

Prerequisites

Basic understanding of Large Language Models (LLMs)
Familiarity with the concept of AI Agents
Basic knowledge of vector databases and retrieval-augmented generation (RAG)

Key Terms

LLM-based Agent: An AI system that uses an LLM as its core controller to perceive, reason, and act in an environment

Reflexion: A specific memory mechanism where an agent critiques its own past actions to generate improved plans for future trials

RAG: Retrieval-Augmented Generation—AI systems that answer questions by first searching for relevant documents

AGI: Artificial General Intelligence—hypothetical intelligent systems with the ability to understand, learn, and apply knowledge across a wide variety of tasks

Parametric Memory: Memory stored implicitly within the model's weights (neural network parameters), updated via fine-tuning or training

Non-Parametric Memory: Memory stored explicitly in external structures (e.g., databases, logs) that the model accesses via retrieval

In-context Learning: The ability of an LLM to learn from examples or information provided in its current prompt/context window without weight updates