Acontext is a skill memory layer for AI agents that automatically captures learnings from agent runs and stores them as readable, editable agent skill files.
The tool addresses the problem of opaque and complicated agent memory by representing all agent knowledge as plain Markdown files rather than embeddings or vector stores. When an agent encounters new information or learns from mistakes, Acontext captures these learnings and evolves existing skill files over time. The agent retrieves skills through explicit tool use and reasoning rather than semantic search, giving developers full visibility into what the agent knows and why it makes decisions.
Acontext suits teams building agents that need to learn across runs while maintaining interpretability and portability. The skill files are framework-agnostic Markdown that work with LangGraph, Claude, AI SDK, or any system that reads files, with no vendor lock-in or re-embedding required when moving skills between agents or LLMs. Developers can version control skills with Git, inspect and edit them directly, and export them as ZIP for reuse in other contexts. The tool lets you define your own memory structure by attaching skills that establish naming conventions and file layouts, rather than imposing a fixed schema.
The project maintains active test coverage across core, API, and CLI components. Development activity shows consistent attention to reliability across multiple deployment surfaces, with the codebase available in both JavaScript and Python ecosystems.
