Treating AI agent skills as long-lived, testable assets with persistent memory—rather than disposable code—significantly improves task success rates and enables skills to transfer between agents and tasks.
This paper introduces MUSE-Autoskill, a framework that helps AI agents continuously improve by creating, storing, and refining reusable skills over time. Instead of treating skills as one-time solutions, the system manages them like software—organizing them in memory, testing them, and learning from experience to make them more reliable and effective across different tasks.