Open-source setup lets Claude, Codex, and Grok all read the same shared memory
When you run multiple AI agents, each one forgets everything once the conversation ends — they have no shared memory. This setup fixes that by using an (a folder of plain text Markdown files) as the single source of truth. A local tool called gbrain indexes those files every five minutes and builds a , then exposes a single MCP endpoint so that Claude, Codex, Grok, and other agents can all query the same information.
Change a fact once in Obsidian and every agent instantly sees the update. Everything runs on your own machine at zero cost per query. The core indexing and graph engine is Garry Tan's gbrain project; OneBrain is the wiring on top — launchd background services, a sync loop, offline embeddings (converting text into searchable number vectors), secret management, and .
To get started, install Obsidian, move the vault out of iCloud into a regular local folder, and put it under git .
Key points
- One acts as shared memory for all your AI agents — edit once, all agents see it
- gbrain auto-indexes the vault every 5 minutes and serves it via a local MCP endpoint
- Claude, Codex, Grok, and other agents can all connect to the same endpoint simultaneously
- Everything — no API calls for retrieval, no cost per query, no data leaving your machine
- Start by moving your to a local git-tracked folder outside iCloud