MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution

MemMA proposes a multi-agent-based memory cycle coordination framework aimed at addressing memory challenges in long-distance interactions. The framework includes a forward path and a backward path: in the forward path, Meta-Thinker generates structured guidance to guide Memory Manager in building knowledge and command Query Reasoner to perform iterative retrieval; in the backward path, an in-situ self-evolution mechanism is introduced to optimize the system by synthesizing probing question-answer pairs, verifying current memories, and transforming failures into repair actions. Experiments show that MemMA outperforms existing baselines on the LoCoMo benchmark, supports various LLM backends and storage architectures, and the related code is publicly available.