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

Analysis

Why This Paper Matters

Large language model (LLM)-based agents have emerged as a promising paradigm for solving complex, real-world problems that require long-term interactions with dynamic environments. Unlike standalone LLMs, agents possess a self-evolving capability that hinges on their ability to remember past experiences and use that information to adapt. However, research on memory mechanisms for LLM-based agents has been fragmented across numerous papers, with no systematic review to distill common design patterns or evaluate trade-offs. This survey fills that critical gap by providing the first comprehensive overview of memory in LLM-based agents, offering a holistic perspective that is essential for both newcomers and seasoned researchers.

The paper's significance is further amplified by the rapid growth of the field. With hundreds of citations already, it is clear that the community has been awaiting such a synthesis. By organizing existing knowledge, the survey enables researchers to build upon a shared understanding rather than reinventing solutions. Moreover, it highlights the importance of memory as a cornerstone for agent-environment interactions, paving the way for more sophisticated and autonomous agents.

Technical Contributions

• Systematic Review: The survey systematically categorizes and compares memory mechanisms proposed in prior work, extracting common designing patterns and effective strategies.
• Conceptual Clarification: It addresses fundamental questions: what memory is in the context of LLM-based agents and why it is necessary for self-evolution and long-term interactions.
• Design and Evaluation Framework: The paper reviews how to design memory modules and how to evaluate their performance, providing a structured approach for future research.
• Application Landscape: It presents various agent applications where memory plays a pivotal role, demonstrating the practical relevance of memory mechanisms.
• Future Directions: By analyzing limitations of existing work, the survey outlines promising research directions, including the need for more scalable, efficient, and adaptive memory systems.
• Living Repository: The authors maintain a GitHub repository to keep track of the latest advances, ensuring the survey remains a valuable resource over time.

Results

As a survey paper, it does not present new experimental results but instead synthesizes findings from a large body of literature. It identifies that memory mechanisms can be broadly categorized into short-term and long-term memory, with various implementations such as vector databases, knowledge graphs, and recurrent architectures. The survey notes that while many promising approaches exist, there is a lack of standardized evaluation benchmarks and metrics, making it difficult to compare different memory designs directly. It also highlights that most existing work focuses on specific application domains, such as dialogue systems or task-oriented agents, and that general-purpose memory mechanisms remain an open challenge.

Significance

This survey is poised to become a foundational reference for the growing community of researchers and practitioners working on LLM-based agents. By providing a clear taxonomy and identifying open problems, it accelerates progress toward more capable, self-evolving agents that can operate effectively in complex, real-world environments. The maintained repository ensures that the survey remains current, fostering ongoing collaboration and innovation. Ultimately, this work contributes to the broader goal of building AI systems that can learn and adapt over time, bringing us closer to truly autonomous agents.