LazyMem: Efficient Long-Term Memory for LLM Agents via Query-Time Construction
A new technique named LazyMem has been developed by researchers to enhance long-term memory in large language model (LLM) agents by postponing memory creation until a query is made. This strategy addresses the challenge of balancing retrieval scope and noise by analyzing candidate pools in overlapping parallel windows with a lightweight model, which selectively stores and compresses only the information pertinent to the query. The model undergoes supervised fine-tuning followed by reinforcement learning, utilizing a reward system that promotes the recognition of relevant messages and the creation of compressed representations. The goal of this method is to boost both efficiency and accuracy in utilizing previous interactions for LLM agents.
Key facts
- LazyMem defers memory construction to query time.
- It processes retrieved candidate pools in overlapping parallel windows.
- A lightweight model selectively retains and compresses query-relevant content.
- Training uses supervised fine-tuning followed by reinforcement learning.
- The reward jointly encourages identification of relevant messages and generation of compressed representations.
- The approach addresses the tension between retrieval breadth and noise.
- It aims to improve efficiency and accuracy for LLM agents.
- The research is presented in arXiv:2607.22690.
Entities
Institutions
- arXiv