This repository provides a comprehensive survey on the memory mechanisms of Large Language Model (LLM) based agents, targeting researchers and developers in the AI field. It systematically reviews and categorizes existing memory solutions, offering insights into their design, evaluation, and application, thereby aiming to inspire future advancements in agent capabilities.

How It Works

The survey explores memory in LLM-based agents through three lenses: cognitive psychology, self-evolution, and agent applications. It posits that mimicking human memory structures provides a cognitive foundation, while memory's role in experience accumulation, environment exploration, and knowledge abstraction is crucial for agent self-evolution. Furthermore, it highlights memory's indispensability in practical applications like conversational agents and role-playing simulations.

Quick Start & Requirements

This is a survey paper, not a software library. No installation or execution is required.

Highlighted Details

• Paper accepted by ACM Transactions on Information Systems (TOIS).
• Provides a systematic review of memory mechanisms for LLM-based agents.
• Discusses "what is" and "why we need" memory in agents.
• Covers implementation strategies and evaluation methods for memory modules.

Maintenance & Community

The paper was released on arXiv on April 21, 2024, and accepted by ACM TOIS on July 2, 2025. Contact information for questions and suggestions is provided via email.

Licensing & Compatibility

This repository contains a survey paper, not code. Licensing information is not applicable.

Limitations & Caveats

As a survey, it summarizes existing work and identifies future directions. It does not introduce new implementations or codebases.