New AI Fixes Software Bugs by Remembering the Right Past Fixes

Software companies now use AI to fix bugs automatically. Think of a huge project's code as a giant filing cabinet. When something breaks, the AI writes a fix. Some AI tools keep a memory of past fixes, hoping that remembering old solutions helps them solve new problems faster.

But the researchers found memory is messy. Some projects have thousands of saved fixes, while others have almost none, so the AI gets little help where it is needed most. They also discovered that more memory is not better. A huge pile of old fixes just adds noise, like a mechanic drowning in repair manuals instead of grabbing the right one. And different repair steps need different memories: finding the cause of a bug is not the same as writing or testing the fix.

So they built a smarter system. When a project has few past fixes, it borrows from similar projects elsewhere. It ranks each memory by how relevant, useful, specific, and non-repetitive it is. And it sorts memories by repair stage, so the AI uses the right kind of experience for reproducing the bug, locating the cause, writing the fix, polishing it, and checking it works.

They tested it on SWE-Bench-Lite and SWE-Bench-Verified, two well-known sets of real bugs from real open-source projects that AI must fix on its own. The system performed better on projects with thin repair history, pulled in less useless memory, and was better at rescuing failed fixes and turning them into working ones. The catch: this is a research paper, not a product you can buy. Its results come from benchmark tests, not your company's tangled code, and it has not yet been reviewed by other scientists.