Memento-Skills introduces an evolving external memory for AI agents, enabling skill development and updates without retraining the base language model.

The team released the code on GitHub and lays out enterprise adoption considerations, emphasizing task structure, governance, and security, with structured evaluation guides and a warning against over-deployment in unsuitable domains.

Benchmark results show GAIA accuracy improving from 52.3% to 66.0% and HLE performance more than doubling, with end-to-end task success using a specialized router at 80% versus 50% with BM25 baselines.

Future work points to multi-agent coordination for longer-horizon tasks and addressing governance and safety for autonomous self-improvement.

A skill router selects the most behaviorally relevant skill, and the system updates or rewrites skills based on execution feedback, using automated unit-test gates to prevent regressions.

Continual learning is driven by one-step offline reinforcement learning, prioritizing long-term utility over superficial text similarity.

The system expanded autonomously from five seed skills to 41 on GAIA and 235 on HLE, showing scalable, self-generated skill growth.

Skills are stored as structured markdown artifacts containing declarative specifications, prompts, and executable code, enabling a Read-Write Reflective Learning cycle.

In production, Memento-Skills avoids heavy data requirements and model fine-tuning, reducing operational overhead typical of traditional skill creation.