Reproduce “LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory” 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
Jing Yu, Yibo Zhao, Jiaming Zhang, Xiang Li
LAZYMEM is a retrieve-then-construct memory system for LLM agents that addresses the challenge of balancing broad information recall with manageable context burden. It uses a lightweight, RL-trained model to selectively construct query-conditioned memory contexts, achieving an LLM-judge accuracy of 0.85 with an average of 213 answer-context tokens while reducing latency compared to prior methods.
Key insights
- Deferring memory construction to query-time, conditioned on the specific query, enables dynamic information retention and compression, proving more effective than query-agnostic pre-processing.
- A lightweight, specialized memory-processing model, refined through a two-phase training approach involving supervised fine-tuning and reinforcement learning, achieves high-quality selective memory construction efficiently.
- Processing broadly retrieved memory through small, overlapping windows in parallel effectively mitigates "context rot" for the construction model and ensures bounded overall latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.