A new research paper from Alibaba, Shanghai Jiao Tong University, and USTC introduces CoEvo-Mem, a framework designed to improve long-term memory in LLM agents. The core insight is that retrieval policy and memory bank are interdependent, yet most existing methods optimize one while keeping the other static. CoEvo-Mem proposes a closed-loop system where both components evolve together, while the underlying LLM remains frozen. This co-evolution allows the agent to adapt its memory strategy based on task demands and accumulated experience. The work is currently available on arXiv and targets a fundamental challenge in building agents that can operate over extended periods. For developers and researchers working on agent architectures, this represents a promising direction to explore, potentially leading to more robust and adaptive memory systems.
CoEvo-Mem, a collaboration between Alibaba, SJTU, and USTC, proposes a closed-loop framework where retrieval policy and memory bank co-evolve for long-term LLM agents. This addresses a key limitation in prior work that optimizes one component while keeping the other fixed. The approach is significant for advancing agent memory systems and improving long-horizon task performance.