EvoDisco: Strategy Discovery for Evolving Agent Memory
Summary
- Researchers introduce EvoDisco, a framework that formulates LLM agent memory evolution as evolutionary strategy discovery.
- EvoDisco uses a learnable policy to select, instantiate, and time semantic operators including Abstract, Refine, Bridge, and Reframe.
- The policy is optimized via a bi-level fitness signal combining inner retrieval-quality improvements and outer strategic efficacy.
- Evaluated on ALFWorld, WebShop, and LifelongAgentBench, EvoDisco outperforms strong memory baselines like MemP and MemRL.
Why it matters
EvoDisco overcomes the limitation of static operator schedules in LLM agent memory management, enabling adaptive and strategic memory evolution that significantly improves agent performance across complex environments.