Reproduce “RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents” 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
Zijie Dai, Shiyuan Deng, Sheng Guan, Yizhou Tian, Xin Yao, Xiao Yan, James Cheng
RecMem introduces a three-tier memory architecture for long-running LLM agents, employing a recurrence-based consolidation mechanism that significantly reduces token consumption for memory construction. This system achieves comparable or improved question-answering accuracy, particularly in temporal reasoning, by selectively processing semantically recurring interactions.
Key insights
- Adopting a recurrence-based, selective memory consolidation strategy drastically reduces the computational cost of building long-term memory for LLM agents.
- Cognitive science theories, such as the multi-store theory and Complementary Learning Systems, provide effective inspiration for designing efficient LLM memory systems.
- A multi-tiered memory architecture allows for robust information management, balancing the preservation of raw details with the generation of coherent event narratives and fine-grained facts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.