altk-evolve

Evolve is an open-source system that gives AI agents persistent memory so they can learn from past sessions instead of starting fresh each time. It targets coding agents and integrates with assistants such as Claude Code, Codex, IBM Bob, and Hermes through an MCP server. The project captures conversation trajectories, extracts reusable guidelines, stores them in a vector backend, and applies LLM-based conflict resolution to update knowledge when new information contradicts existing guidelines. The Lite version is designed to slot into existing agent workflows, while the full MCP server exposes tools for saving trajectories, creating and retrieving entities, publishing or unpublishing guidelines across namespaces, managing retention policies, and checking compliance status. Storage backends include a default filesystem option with text matching, PostgreSQL with pgvector for production use, and Milvus for large-scale vector search. A web UI is available for inspecting and managing the knowledge base. Evolve reports benchmark improvements on AppWorld, with a +8.9 point overall reliability gain and a 74% relative increase on hard multi-step tasks. The project is backed by research papers on trajectory-informed memory generation and agent consistency, and it is released under the Apache 2.0 license.