Acontext

Acontext is an open-source memory layer for AI agents that captures learnings from agent sessions and stores them as plain Markdown skill files, which can be read, edited, versioned with Git, and shared across agents, frameworks, and LLMs. Rather than relying on opaque vector embeddings or semantic search, it uses a distillation pipeline that processes session messages and task outcomes through a Skill Agent, which routes extracted knowledge into structured files defined by a user-supplied `SKILL.md` schema. Agents retrieve stored skills at runtime by calling `get_skill` and `get_skill_file` tools directly, letting reasoning drive retrieval instead of top-k similarity. The system integrates with Claude Code via a one-line prompt pointing to the hosted `SKILL.md`, and works with any LLM or framework that can read files, including LangGraph and Vercel AI SDK. Skills can be downloaded as a ZIP archive with no re-embedding required, and a self-hostable backend can be spun up locally using `acontext-cli` and Docker. Teams building persistent, self-improving agents who want inspectable and editable memory rather than black-box retrieval stores are the primary audience.

3 items in this repository

What people ask about Acontext