RESEARCH · RESEARCH · #3798
EvoDisco: evolutionary strategy discovery to evolve agent memory
EvoDisco (arXiv:2611.22102v1) frames agent memory evolution as an evolutionary strategy discovery problem: a learned policy selects and instantiates one of four semantic operators—Abstract, Refine, Bridge, Reframe—rather than applying a fixed memory-update schedule. Using a bi-level fitness (inner retrieval-quality regularized by complexity; outer strategic efficacy over past decisions), EvoDisco outperforms strong memory baselines MemP and MemRL on ALFWorld, WebShop, and LifelongAgentBench, with the largest gains on ALFWorld (+8.8 online / +12.9 frozen-transfer vs MemP; +16.4 / +17.9 vs MemRL).
KEY POINTS
- EvoDisco (arXiv:2611.22102v1) frames agent memory evolution as an evolutionary strategy discovery problem: a learned policy selects and instantiates one of four semantic operators—Abstract, Refine, Bridge, Reframe—rather than applying a fixed memory-update schedule.
- Using a bi-level fitness (inner retrieval-quality regularized by complexity; outer strategic efficacy over past decisions), EvoDisco outperforms strong memory baselines MemP and MemRL on ALFWorld, WebShop, and LifelongAgentBench, with the largest gains on ALFWorld (+8.8 online / +12.9 frozen-transfer vs MemP; +16.4 / +17.9 vs MemRL).
- It replaces fixed memory-update rules with learned, multi-operator strategies and shows measurable gains on standard agent benchmarks, advancing lifelong-memory methods for language-model agents.
WHY IT MATTERS
It replaces fixed memory-update rules with learned, multi-operator strategies and shows measurable gains on standard agent benchmarks, advancing lifelong-memory methods for language-model agents.