‹ back to 2026-09-20

AdaRepair-Mem: Adaptive Experience Orchestration for Repository-Level Program Repair

arXiv cs.SE LLM Agents Z. C. Luo, J. C. Guo, W. J. He, S. Y. Wang, J. C. Yu, F. M. Zhao, Y. Chen, T. Cao, L. Q. Liu, N. Zheng, W. Xu, J. Jiang, Z. M. Zhao 2026-09-17

TL;DR - AdaRepair-Mem is an adaptive memory-retrieval framework for LLM-based repository-level program repair. It improves repair by selecting relevant, high-quality experiences for each repair stage rather than simply accumulating more memories.

Coverage-aware retrieval uses cross-repository or repair-type memories when repository-specific examples are scarce.
Quality-aware selection ranks experiences by relevance, historical utility, specificity, and redundancy.
Stage-aware routing separates memories for reproduction, localization, patch generation, refinement, and validation.
Evaluations on SWE-Bench-Lite and SWE-Bench-Verified show gains for under-covered repositories, less noisy retrieval, and better failed-to-fixed patch refinement.

view merged work →