Acontext: a transparent skill memory layer that stores agent learnings as editable Markdown files

What it solves

Acontext provides a transparent, human-readable memory layer for AI agents. It solves the problem of "opaque" agent memory—where learnings are hidden in complex embeddings or proprietary databases—by storing agent skills and knowledge as simple Markdown files that can be read, edited, and shared across different LLMs and frameworks.

How it works

Acontext operates through a cycle of distillation and recall:

1. Store (Learning): The system monitors session messages. When a task is completed or fails, an LLM-driven distillation process analyzes the execution trace to determine what worked and what didn't. A "Skill Agent" then writes these learnings into Markdown files based on a user-defined schema.
2. Recall (Usage): Instead of using semantic search (top-k), agents use specific tools (get_skill and get_skill_file) to explicitly fetch the knowledge they need. This allows the agent to reason about what information to retrieve, providing a process of progressive disclosure.

Who it’s for

Developers building AI agents who want their agents to learn from mistakes and reuse successful strategies without being locked into a specific vendor, database, or framework (e.g., LangGraph, Claude, or Vercel AI SDK).

Highlights

• Human-Readable Memory: All memories are stored as Markdown files, making them easy to debug, version control with Git, and inspect.
• Framework Agnostic: Works with any framework that can read files, avoiding API lock-in.
• Agent-in-the-Loop Retrieval: Uses tool-based retrieval rather than opaque embedding search.
• Portable Knowledge: Skills can be exported as ZIP files and moved between different agents or LLMs.
• Integrated Tooling: Includes a virtual persistent filesystem (Disk), isolated code execution (Sandbox), and SDKs for Python and TypeScript.