# Acontext

An open-source skill-based memory layer for AI agents that automatically captures what an agent learns from its runs and stores it as readable, editable agent skill files shared across agents and frameworks.

Acontext is an open-source memory layer for AI agents that represents memory as agent skill files rather than an opaque store, so what an agent has learned can be read, edited and shared like any other skill. It automatically captures learnings from an agent's runs and writes them out in the same skill format the agent already uses.

Memories are plain Markdown files that work with any framework that can read files, such as LangGraph, Claude or the AI SDK, without embeddings or vendor lock-in, and a user can attach additional skills to define the schema, naming and layout of a given memory type. An agent retrieves what it needs through explicit tool calls that fetch a skill or a skill file, rather than semantic top-k search, and memory files can be exported as a ZIP to run locally, in another agent, or with another LLM.

Its process for turning a completed or failed task into an updated skill runs distillation through a dedicated skill agent.

Website: https://acontext.io