Reproduce “MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minhua Lin, Zhiwei Zhang, Hanqing Lu, Hui Liu, Xianfeng Tang, Qi He, Xiang Zhang, Suhang Wang
The Pennsylvania State University, Amazon, and Microsoft researchers developed MEMMA (Memory Cycle Multi-Agent Coordination), a multi-agent framework that enhances large language model (LLM) agents' long-term memory by coordinating memory construction, retrieval, and utilization processes. MEMMA achieved an 81.58% overall accuracy on the LoCoMo benchmark, improving upon the strongest baseline by 5.92%, and demonstrated strong generalization to other benchmarks.
Key insights
- Explicitly coordinating the memory cycle through a multi-agent framework with both forward (strategic guidance) and backward (self-evolutionary repair) paths significantly improves LLM agent performance in long-term memory tasks.
- A Meta-Thinker providing strategic guidance for memory construction and iterative retrieval effectively addresses myopic decisions and aimless searches, leading to more accurate and coherent memory.
- Implementing an "in-situ" self-evolution mechanism for immediate memory verification and repair during construction effectively addresses sparse and delayed supervision, producing a more robust and correct memory bank.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.