Memento-Skills

Why we included this project

Most agent frameworks treat tools as a fixed list you set once and forget. Memento-Skills instead keeps capabilities in a skill library the agent can pull up, run, judge, and rewrite over time, cycling through Read-Execute-Reflect-Write. When a step fails, the agent treats that as a signal to patch the weak skill and write the improved version back into memory, so it gets better from live use instead of just retrying. It ships with a CLI, a desktop GUI, and a Feishu bridge, and runs locally with persistent state, which makes it more than a research demo. For teams that want an agent to learn from real tasks without retraining the underlying model, this deployment-time learning is a practical alternative to fine-tuning.