# Acontext

> Open-source skill memory layer that captures learnings from AI agent runs and stores them as readable Markdown skill files.

Acontext is an open-source skill memory layer for AI agents, offered as both a managed cloud platform and a self-hostable stack. It captures learnings from agent sessions and writes them as Markdown agent skill files that can be read, edited, and shared across agents, LLMs, and frameworks. Python and TypeScript SDKs are available.

## What It Is

Acontext sits alongside an agent and turns what the agent did into reusable files. Instead of opaque embeddings, memory is stored in the agent skills format (SKILL.md schema), so developers can see what the memory contains and correct it. Beyond skill memory, it also provides session message storage, disk and artifact storage, sandboxed code execution, context editing, and an observability dashboard.

## How Skill Memory Works

Session messages are the raw input, and tasks are extracted from the message stream automatically or from explicit outcome reporting. When a task completes or fails, an LLM distillation pass infers what worked, what failed, and user preferences. A skill agent then decides whether to update an existing skill or create a new one, writing according to the developer's SKILL.md schema. On the next run, the agent calls get_skill and get_skill_file tools to fetch what it needs, with retrieval by tool use rather than semantic top-k search.

## Storage, Context and Observability

The product describes short-term memory for messages in OpenAI, Anthropic, and Gemini formats with automatic conversion, S3-backed disk storage with glob and regex search, and isolated sandboxes with mountable skills. Context engineering features include context editing without altering stored messages, prompt cache stability, and session summaries. The dashboard shows agent tasks, traces, and success rates.

## Setup Path

Users can get an API key from the cloud dashboard and run pip install acontext, or install the TypeScript package. For self-hosting, the acontext-cli runs the backend through Docker with an OpenAI API key, exposing an API on port 8029 and a dashboard on port 3000. Skills can be downloaded as ZIP files and reused in other agents or LLMs. Templates cover OpenAI, Claude Agent SDK, Agno, smolagents, and Vercel AI SDK. The repository is licensed under Apache 2.0.

## Features
- Automatic skill distillation from agent sessions into Markdown files
- Learning spaces
- Agent skill tools get_skill and get_skill_file
- Skill export as ZIP
- Multi-format message storage (OpenAI, Anthropic, Gemini)
- Disk and artifact storage with glob and regex search
- Sandbox code execution with mountable skills
- Context editing and session summaries
- Prompt cache stability
- Task extraction, traces, and success-rate dashboard
- Python and TypeScript SDKs
- Self-hosting via acontext-cli
- Bearer token API authentication

## Integrations
OpenAI, Anthropic, Gemini, LangGraph, Agno, Claude Code, OpenClaw, Claude Agent SDK, smolagents, Vercel AI SDK

## Platforms
WEB, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
sdk-ts/v0.1.21

## Links
- Website: https://acontext.io
- Documentation: https://docs.acontext.io/
- Repository: https://github.com/memodb-io/Acontext
- EveryDev.ai: https://www.everydev.ai/tools/acontext