# osk-system

A source-grounded MCP memory runtime for LLM agents that stores knowledge as local Markdown files in an Obsidian-compatible vault, keeping a person's approved knowledge separate from what an agent has only proposed.

osk-system keeps an AI agent's memory as local Markdown files rather than a hidden database, linking each piece of knowledge back to its source and to the specific conversation round it came from, and it separates a protected region a person has approved from changesets an agent has only proposed for review. A Python engine exposes MCP search and reading tools over stdio and stores knowledge with BM25 retrieval; Obsidian is an optional way to browse the resulting graph, and Git synchronization of the vault is opt-in.

Setup can be handed to a supported agent, currently Codex, Claude Code, Kiro or Antigravity, by pasting a one-line install instruction that has it read a remote install guide, ask where to create the vault and which optional features to enable, and show its plan before applying it.

It is described by its maintainer as a developer public beta used daily on Windows 11, with CI running its test suite on Windows and Linux and, as a non-blocking job, macOS; a vault is an ordinary Git repository that the project recommends keeping backed up. It is written in Python.

Website: https://github.com/lpaiu-cs/osk-system