Evolver
A self-evolving engine for AI agents that uses a Genes-Capsules-Events model (GEP) to make an agent's learned skills and memory auditable as it evolves.
Evolver is a self-evolving engine for AI agents, built around what it calls GEP: Genes, Capsules and Events. It gives an agent a memory and skill-library system that evolves as the agent works, with each step of that evolution kept auditable.
A companion paper compares two representations for what an agent learns from experience: documentation-style Skill packages, which the paper reports give an unstable and sparse training signal, against a more compact Gene representation, which performs better overall and stays more robust under structural change, across thousands of trials on scientific code-solving tasks.
It installs as an npm package or by cloning the repository, and is written in JavaScript for Node.js. Already-published versions remain available under their original MIT or GPL-3.0-or-later terms; the project has stated that future releases will move to a source-available model rather than being fully open source.