Program repair improves when memories match repository needs and repair stages

Memory-augmented program repair uses past repair experiences to help LLMs fix bugs, but existing approaches have three problems: (1) memory is unevenly distributed across repositories, leaving low-resource repos unsupported; (2) more memories do not always help; relevance and quality matter more; (3) repositories often have abundant reproduction memories but scarce patch or refinement memories. The authors propose AdaRepair-Mem, which addresses these with three mechanisms: coverage-aware retrieval (uses cross-repo memories when same-repo is thin), quality-aware selection (ranks by relevance, utility, specificity, redundancy), and stage-aware routing (retrieves different memories for each repair stage). Evaluated on SWE-Bench-Lite and SWE-Bench-Verified, the framework improves repair on under-covered repositories, reduces noisy retrievals, and better converts failed patches to successful fixes. This is interesting because it shows that careful orchestration of memory retrieval matters more than simply accumulating historical data.