$\varepsilon$-MemEvo Boosts LLM Program Evolution with Cross-Task Memory

Existing LLM-based program evolution systems, like FunSearch, typically optimize each task in isolation, discarding valuable search experience after completion. This new research introduces $\varepsilon$-MemEvo, a framework designed to facilitate cross-task knowledge transfer in LLM program evolution. Instead of storing raw code, $\varepsilon$-MemEvo captures prior experience as compact, task-agnostic natural-language summaries of successful algorithmic strategies, making them transferable across diverse tasks with varying APIs and evaluators. To mitigate the risk of negative transfer from semantically mismatched memories, $\varepsilon$-MemEvo incorporates an adaptive injection gate. This gate intelligently decides whether and at what intensity retrieved memories should be applied. Evaluation across eight diverse optimization benchmarks, including mathematical optimization and systems engineering, demonstrated significant improvements. Using a GPT-5 backbone, $\varepsilon$-MemEvo achieved a mean relative gain of +8.7% in Area Under the Curve of Convergence (AUCC) over AdaEvolve and improved early-stage convergence by +9.4% on average, with less than 1% computational overhead. Ablation studies confirmed the necessity of adaptive gating, as naive memory injection could lead to catastrophic failures.