For agent repair memory, stage-aware routing beats volume: more retrieved experience does not monotonically improve fixes

arXiv 2609.20130·medium signalXBlueskyLinkedInCopy link

AdaRepair-Mem diagnoses three failures in repository-level memory retrieval that any agent-memory builder will recognize: episodic memory is badly imbalanced across repositories so low-resource repos get nothing useful, more memory does not monotonically raise success because relevance and redundancy dominate raw volume, and accumulation is phase-misaligned, with piles of reproduction traces and almost no patch or refinement ones. The fix is three mechanisms: coverage-aware retrieval that falls back to cross-repository or repair-type memory, quality-aware selection ranking on relevance, historical utility, specificity and redundancy, and stage-aware routing that keeps separate memory pools for reproduction, localization, patch generation, patch refinement and validation. On SWE-Bench-Lite and SWE-Bench-Verified it improves under-covered repositories and failed-to-fixed refinement specifically.