Memory in LLM-Based Multi-agent Systems: Mechanisms, Challenges, and Collective Intelligence

Memory plays a central role in transforming Large Language Model (LLM)-based agents from reactive predictors into consistent, context-aware collaborators. While LLM-based single-agent memory has been extensively studied, memory in LLM-based Multi-Agent Systems (LLM-MAS) lacks a systematic taxonomy and review. In multi-agent contexts, memory becomes a shared cognitive infrastructure enabling collective intelligence, long-term coordination, and team evolvement. This survey provides the first comprehensive review of memory in LLM-MAS, synthesizing research across memory architectures, management and operations, evaluation, and application, while formalizing key definitions and introducing the design space. Our survey reveals that memory in LLM-MAS is not a trivial extension of single-agent memory but a distinct research frontier—with new challenges in synchronization, access control, scalability, alignment, and safety. By consolidating diverse literature, we aim to provide a foundation for future research in building intelligent, memory-augmented multi-agent systems. In order to follow the latest works in this field, we create and maintain a repository at https://github.com/ShanglinWu/LLM-MAS_Memory_Survey .