AI Agents Learn to Evolve Skills Autonomously

Researchers have introduced Memento-Skills, a novel framework designed to equip autonomous AI agents with self-evolving capabilities. This innovative system addresses the challenge of adapting AI agents to changing environments without necessitating the retraining of underlying large language models (LLMs).

Memento-Skills functions as a dynamic external memory, allowing agents to continually enhance their functionalities. The framework provides a set of skills that are updated and expanded based on environmental feedback, bypassing the operational overhead associated with fine-tuning model weights or manual skill development.

The system's 'Read-Write Reflective Learning' mechanism enables continual learning by treating memory updates as active policy iteration. When faced with a new task, Memento-Skills retrieves the most behaviorally relevant skill, executes it, and then reflects on the outcome to refine or create new skills, significantly boosting end-to-end task success rates.

Evaluated on rigorous benchmarks like GAIA and Humanity's Last Exam, Memento-Skills demonstrated substantial performance improvements. Powered by Gemini-3.1-Flash, the framework autonomously scaled its initial five seed skills to libraries of 41 and 235 skills, respectively, achieving significantly higher accuracy compared to static skill libraries and semantic retrieval baselines.

While the code is publicly available, enterprise adoption hinges on domain alignment, particularly for structured workflows where skill composition and improvement are key. The researchers caution against deploying Memento-Skills in areas like physical agents or tasks with very long horizons, emphasizing the need for guided self-development and robust evaluation systems for reliable self-improvement.