ALTK Evolve

ALTK Evolve gives AI agents persistent memory of earlier tasks so they can reuse lessons across sessions. It's for coding assistant users and agent builders who want guidance drawn from past successes and mistakes.

• Learning from past work: Evolve analyzes agent conversations and action histories, turns them into reusable guidelines, and retrieves relevant guidance for later tasks.
• Conflicting advice: LLM-based conflict resolution updates existing guidelines when new findings contradict them.
• Memory search: The default filesystem backend finds guidance through text matching; optional Milvus and Milvus Lite backends support search by meaning.
• Trace integration: Phoenix integration imports recorded agent activity, and guideline provenance links stored advice to its originating task or trace.
• Selective sharing: Stored entries are private by default, and agents can publish individual entries for retrieval across namespaces through MCP tools.

Evolve Lite uses file-based memory within assistants including Claude Code, Codex, IBM Bob and Hermes. Full mode runs as a self-hosted MCP server with a local web interface. Sharing controls are available through MCP tools, while the corresponding web interface controls and REST routes aren't implemented.

The Python project is open source under Apache 2.0. Full mode supports direct OpenAI access and model access through a LiteLLM proxy. Self-hosting the memory service doesn't make model processing local: content sent to an external model provider leaves your server.