# Acontext: Your AI Agent's Memory, Saved in Plain Markdown

Ever watched an AI agent repeat the same mistake twice? Or struggle to apply something it just learned to a new task? Most agent frameworks store internal state in opaque JSON blobs or databases, making it hard to inspect, edit, or share what the agent actually "knows."

[**Acontext**](https://github.com/memodb-io/Acontext) flips the script. It's an open source skill memory layer that stores everything an agent learns as plain Markdown files. No black box. No proprietary format. Just files you can read, write, and version control like any other code.

## What It Does

Acontext acts as a persistent memory for AI agents. When an agent learns a new skill – say, how to format a date or query an API – Acontext saves that knowledge as a Markdown file in a structured directory. Later, when the agent encounters a similar task, it can pull relevant skills from those files and apply them.

Think of it as a portable, human-readable knowledge base for your agents. Each skill is a self-contained Markdown doc with a clear heading, a description, and example usage. No hidden state, no embeddings you can't see.

## Why It's Cool

**1. Human readable, version controllable.** Markdown is the lingua franca of developers. You can diff skills, open PRs to improve them, or even write new ones by hand. This makes agent behavior auditable and collaborative in ways that stored embeddings or serialized objects never are.

**2. Modular and composable.** Skills are stored per agent, per project, or globally. You can mix and match skills from different sources, or treat the whole folder as a library. If an agent learns something valuable, you can copy that Markdown file to another agent's memory directory.

**3. No vendor lock in.** Because it's just Markdown, you own your agent's knowledge. You can move it between frameworks, edit it in any text editor, or even export it to other systems. The only dependency is a filesystem.

**4. Easy debugging.** When your agent does something weird, you can open the relevant skill file and see exactly what it was taught. No digging through logs trying to reconstruct what went into a vector store.

## How to Try It

Acontext is a Python package. Install it with pip:

```bash
pip install acontext

Then, in your agent code, initialize a memory store:

from acontext import Memory

memory = Memory("./agent_skills")  # path to a folder of .md files
skill = memory.get_skill("date_format")

You can also add new skills on the fly:

memory.add_skill("date_format", 
    description="Format dates as YYYY-MM-DD",
    usage="date_format("2024-03-15")")

Check the full README for more examples and integration tips.

Final Thoughts

Acontext isn't trying to be the most sophisticated memory system in the world. It's trying to be the most sensible one. By storing agent knowledge as plain Markdown, it makes something that's usually opaque and fragile into something solid, inspectable, and portable.

For solo devs building agents, this means fewer "why did it do that" moments. For teams, it means you can treat agent memory like code – with reviews, tests, and a history. That's a small shift that could pay off big the first time you need to fix a bad behavior without retraining or guessing.

Give it a try on your next side project. Your future self (and your agent) might thank you.

Follow @githubprojects for more developer tools and open source projects.