PAPER / ARXIV:2609.20130
Luo, Z.; Guo, J.; He, W.
RESUMO
Recent memory-augmented repository-level program repair methods reuse historical experiences to improve LLM-based issue resolution. However, analysis reveals three limitations in existing repository-level memory retrieval: episodic memory is imbalanced across repositories, more memory does not monotonically lead to higher repair success, and memory accumulation is phase-misaligned. To address these problems, we propose an adaptive experience retrieval framework for program repair introducing coverage-aware retrieval, quality-aware selection, and stage-aware routing. Evaluated on SWE-Bench-Lite and SWE-Bench-Verified, the proposed framework improves repair performance on under-covered repositories and reduces noisy memory retrieval.
NO MESMO MAPA