Intel / article

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

Source: arxiv.org
Author: 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
Date: Published Sep 18, 2026

Context

AdaRepair-Mem addresses a problem with AI coding agents that reuse earlier bug fixes: a larger memory bank can still supply the wrong experience. The researchers report uneven coverage across repositories and repair stages, with some collections rich in bug-reproduction examples but poor in patch-refinement examples. Research abstract.

Their framework retrieves experience appropriate to the current repair stage, ranks it for relevance and usefulness, and draws on other repositories when local examples are scarce. The authors report improved repair performance on under-covered repositories in two evaluations. The useful design principle is to match past experience to the current repair problem and stage, rather than simply adding more stored examples. Method and findings.

Key takeaways · AI-distilled

Analysis of memory-augmented program repair found more stored repair experience doesn't reliably improve success -- relevance, quality, and redundancy of retrieved memories matter more than raw volume.
Memory is also phase-misaligned: repos often accumulate many bug-reproduction experiences but far fewer patch-generation or patch-refinement ones, leaving later repair stages under-supported.
AdaRepair-Mem's coverage-aware retrieval falls back to cross-repository or repair-type memories when a repo's own history is too thin, helping low-resource repositories that otherwise get little support.
On SWE-Bench-Lite and SWE-Bench-Verified, the framework improved repair on under-covered repositories and better supported turning failed patches into fixed ones.

Terms in this piece · Glossary

SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.

Read the source arxiv.org