developer tools
The GUI Agent Kept a Versioned Memory of What Worked
A persistent skill library improved a fixed computer-use stack across four observed domains—but repeated accepted revisions did not always recover a task.
Summary
A persistent skill library improved a fixed computer-use stack across four observed domains—but repeated accepted revisions did not always recover a task.
The framework turns interaction trajectories and evaluator feedback into versioned procedures that become available on later iterations without changing model weights. After a five-iteration empty-library warm-up, the full system's post-warm-up mean evaluator score exceeded a matched empty-library control in all four OSWorld domain runs by 5.7 to 18.6 percentage points. A provenance study in GIMP found skills crossing task-of-origin boundaries, but also revision churn: repeated accepted edits could fail to recover the task that created the skill. The evidence supports conditional, auditable memory gains—not automatic compounding improvement.
Why it matters
A persistent skill library improved a fixed computer-use stack across four observed domains—but repeated accepted revisions did not always recover a task.
Limits and context
- The evidence supports conditional, auditable memory gains—not automatic compounding improvement.
Key claims
A persistent skill library improved a fixed computer-use stack across four observed domains—but repeated accepted revisions did not always recover a task.
Qualification: The evidence supports conditional, auditable memory gains—not automatic compounding improvement.
Evidence: source-2026-09-08-011
Sources
- arXiv preprint 2609.04869arXiv · primary research
Corrections
No corrections have been recorded for this story.