AI/ ai · program-repair · llm-memory · swe-bench

New Framework Retrieves the Right Memory for AI Bug Fixes

A new framework routes AI code-repair memories by repo, quality, and stage, showing more data doesn't mean better bug fixes.

AI systems that fix code by reusing memories of past repairs turn out to have a curation problem, not a storage problem.

Researchers analyzing repository-level program repair found that memory-augmented systems accumulate experience unevenly: some repositories are data-rich, others barely covered. Piling on more memories does not reliably improve repair success, and the experiences that do exist often skew toward one repair phase, like reproducing a bug, while patch generation and refinement go under-served. To fix this, the team built a framework that falls back to cross-repository or repair-type memories when local data is thin, ranks memories by relevance and redundancy rather than volume, and routes retrieval separately for each repair stage: reproduction, localization, patching, refinement, and validation. Tested on SWE-Bench-Lite and SWE-Bench-Verified, it improved results on under-covered repositories and cut down on noisy retrieval.

This matters because most memory-augmented AI pitches, in coding and elsewhere, assume bigger context and more logged experience automatically means better output. This work argues the opposite: relevance and redundancy control matter more than raw volume, and mismatched retrieval actively hurt performance on some repositories.

It is a small, unglamorous fix, but it points at a real limit on the add-more-memory approach that a lot of agentic-coding tools are betting on.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →