Researchers Develop Memento-Skills System for Self-Evolving AI Agents

A team of researchers has developed Memento-Skills, a system that gives AI agents persistent memory and the ability to self-evolve through a mechanism called Read-Write Reflective Learning. The system stores skills as structured markdown files containing declarative specifications, specialized instructions, and executable code. When an agent encounters a task, it retrieves the most behaviorally relevant skill rather than relying on semantic similarity alone. After execution, the system reflects on the outcome and actively mutates its memory. If execution fails, an orchestrator evaluates the trace and rewrites the skill artifacts directly, patching specific failure modes or creating entirely new skills when needed. The skill router updates through one-step offline reinforcement learning that learns from execution feedback rather than text overlap. Automatic unit-test gates prevent regression On the GAIA benchmark, which tests complex multi-step reasoning, Memento-Skills achieved 66.0% accuracy compared to 52.3% for static baseline systems. On Humanity's Last Exam, an expert-level benchmark, the system more than doubled baseline performance from 17.9% to 38.7%. The researchers have released the code on GitHub, noting that the system works best in structured workflow environments where skills can be composed and evaluated.