Reproduce “LeanMem: Simple and Efficient Long-Term Memory for 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
Yuxin Liao, Le Wu, Min Hou, Hao Liu, Han Wu, Zishu Wang
LeanMem, a memory framework for LLM agents, introduces a differentiated approach to long-term memory management, significantly improving answer accuracy while reducing token consumption and latency. It achieves this by handling heterogeneous dialogue content through controlled writing, selective evolution, and adaptive evidence composition, tailoring storage and retrieval strategies to information characteristics.
Key insights
- Conversational information is heterogeneous, requiring differentiated processing, storage, and retrieval strategies tailored to its stability, temporal dependence, and fidelity needs, rather than uniform summarization.
- Employing distinct memory types (profile, event, record) with type-specific maintenance (e.g., only evolving event memory) is crucial for simultaneously improving both efficiency and the preservation of critical information.
- Adaptive retrieval planning, which dynamically selects memory types and allocates resources based on a query's specific evidence requirements, is essential for optimizing inference costs and ensuring the composition of highly relevant evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.