EvolveMem: Self-evolving memory boosts LLM agents by 25.7%

EvolveMem, introduced by Jiaqi Liu and co-authors, tackles a key limitation in long-term memory for LLM agents: existing systems treat retrieval infrastructure as fixed after deployment. While stored content evolves, scoring functions, fusion strategies, and answer-generation policies remain frozen. EvolveMem breaks this pattern by exposing its full retrieval configuration as a structured action space, optimized by an LLM-powered diagnosis module. In each evolution round, the module reads per-question failure logs, identifies root causes, and proposes targeted adjustments. A guarded meta-analyzer applies changes with automatic revert-on-regression and explore-on-stagnation safeguards, creating a closed-loop self-evolution process that replaces manual tuning.

Starting from a minimal baseline, the system autonomously converges on effective retrieval strategies, even discovering entirely new configuration dimensions not present in the original action space. On the LoCoMo benchmark, EvolveMem outperforms the strongest baseline by 25.7% relative and achieves a 78% improvement over the minimal baseline. On MemBench, it exceeds the strongest baseline by 18.9% relative. Notably, evolved configurations transfer positively across benchmarks, suggesting the self-evolution process captures universal retrieval principles rather than benchmark-specific heuristics. The code is publicly available, enabling further research into self-optimizing agent architectures.