A new framework lets AI agents that operate computer interfaces build a shared memory of skills instead of forgetting every lesson after each task.
Researchers describe an online skill-evolution system for computer-use agents - AI that clicks, types, and navigates real software the way a human would. Instead of discarding what an agent learns after each attempt, the framework turns interaction logs and evaluator feedback into a versioned library of reusable procedures. Each run executes against a frozen snapshot of that library, and only evidence-backed updates get folded in for the next round, without touching the underlying model's parameters. The team tested it across four application domains in the OSWorld benchmark, using the same action-generation and GUI-grounding stack for both a skill-building version and a no-memory control.
After a five-iteration warm-up with no skills to draw on, the memory-equipped agent beat the blank-slate version in all four domains, by 5.7 to 18.6 percentage points on average. That is a meaningful gap for a change that costs no retraining - it behaves more like handing the agent a shared notebook than upgrading its brain. It is also a rare study that checks not just whether performance improves, but whether the underlying memory stays trustworthy as it accumulates.
In GIMP, the library was not infallible: repeated edits sometimes got accepted and reused even when they could no longer be traced back to the task that produced them - a reminder that letting an AI agent curate its own memory needs an audit trail, not just a feedback loop.