CoEvo-Mem: A Closed-Loop Framework for Co-Evolving Retrieval and Memory in LLM Agents
There's a new study out that presents CoEvo-Mem, a system designed to improve long-term LLM agents by evolving both their memory banks and retrieval policies at the same time. You can find it on arXiv under the number 2608.01739. The researchers point out that most existing methods either focus on how to access memory better or how to update it, but they often miss the connection between retrieval and memory activation. CoEvo-Mem tackles this by using a static LLM to create specific query rewrites, with a lightweight router making real-time adjustments. The context retrieved serves as a link: task outcomes give feedback on routing, and this, in turn, helps refine the memory bank, enhancing LLM performance across tasks and sessions.
Key facts
- The paper is titled 'CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents'.
- It is available on arXiv with identifier 2608.01739.
- The paper proposes a closed-loop framework for co-evolving retrieval policy and memory bank.
- Existing methods typically optimize either memory access or memory evolution separately.
- CoEvo-Mem uses a frozen LLM to generate route-specific query rewrites and a routing prior.
- A lightweight residual router corrects the routing prior online.
- Retrieved context serves as the coupling interface between retrieval and memory update.
- Task outcomes assign credit to routing decisions, and trajectory-conditioned feedback updates the memory bank.
Entities
Institutions
- arXiv