A team of university researchers has introduced Memento-Skills, a pioneering framework that allows AI agents to autonomously update and expand their skills without needing to retrain their core language models. This groundbreaking approach uses an evolving external memory composed of structured markdown skill files, which the AI refines through a cycle of execution, feedback, and self-modification using a mechanism called Read-Write Reflective Learning. Unlike traditional methods that rely on static skill libraries or simple text-based retrieval, Memento-Skills retrieves and upgrades skills based on actual performance and utility, significantly improving task success rates. In benchmarks such as GAIA and Humanity’s Last Exam, this system demonstrated substantial gains, including a 13.7-point accuracy improvement and more than doubling performance in complex academic tasks. This technology is especially promising for enterprise applications involving structured workflows, as it reduces operational costs linked to retraining and manual skill updates. To ensure stability and security, automated skill updates undergo rigorous testing before implementation. The researchers have made the code available on GitHub, inviting further exploration and adoption in practical settings.
Back
