Reproduce “BMAM: Brain-inspired Multi-Agent Memory Framework” 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
Yang Li, Jiaxiang Liu, Yusong Wang, Yujie Wu, Mingkun Xu
Researchers from Guangdong Institute of Intelligence Science and Technology, Institute of Science Tokyo, and The Hong Kong Polytechnic University developed BMAM, a brain-inspired multi-agent memory framework designed to counteract "soul erosion" in LLM agents. This framework mitigates the degradation of temporal coherence, semantic consistency, and identity preservation, achieving 78.45% accuracy on LoCoMo and a 72.90% personalized response rate with 0.1% inconsistency on PrefEval.
Key insights
- The functional specialization of human memory, with distinct subsystems for different types and timescales of information, provides an effective architectural pattern for robust AI agent memory.
- Different aspects of "soul erosion" (temporal, semantic, identity) can be directly mitigated by mapping them to specialized, interacting memory components (e.g., hippocampus-inspired episodic memory for temporal coherence).
- A dynamic "memory loop" that includes explicit processes for encoding, consolidation, retrieval, and reflection is crucial for allowing agents to adaptively learn and refine their memory over prolonged interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.