The Three-Layer AI Operating System
Nick Milo has 17,000 notes in his Obsidian vault. When he pointed an AI at it and let it do its thing, the result was usable. Not because the AI was good at navigating chaos, but because of a translation layer he built between the vault and the AI.
His framework has three layers. The first: the Ideaverse. Your own thoughts and notes , the raw material. The second: the AIOS folder. A separate area inside the vault that contains maps, manuals, and the MIMD file , a portable identity for the AI that works with any tool, not just Claude. The third: the external AI layer. Claude Cowork, or whatever tool you plug in.
The Problem With CLAUDE.md
If you have used Claude Cowork before, you know about CLAUDE.md , the master file Claude reads every time you start a new session, giving it context about your project and preferences.
The problem: CLAUDE.md is Claude-specific. If you switch to a different AI tool next month, that file does not automatically go with you. It is optimised for one model's conventions.
Milo's alternative is the MIMD file. Same concept , a plain markdown file that establishes context at the start of every session , but written for any AI, not just Claude. When Claude goes away or gets replaced by something better, the MIMD file travels with you. The identity layer is not locked to any vendor.
Why the Separation Matters
The AIOS folder sits inside the vault but is deliberately separated from the Ideaverse. The reason: AI-generated content should be isolatable. If you want to clear out AI outputs and start fresh, you should be able to do that without touching your actual notes and ideas.
This separation preserves the integrity of your thinking. The AI layer generates, synthesises, and organises. The Ideaverse is where your original thought lives. Keeping them separate is what makes the system sustainable over years, not just useful for a few weeks.
The Bet on Local Models
Milo is explicit about where he thinks this goes. Eventually, the external AI layer will probably be a local model running on Apple silicon. The API-based approach is a transitional phase. Building a system that can swap out the AI layer , Claude today, something else next year, local model in 2028 , is the durable bet.
His file format is plain markdown. His MIMD file is not tied to any tool. The maps and manuals in the AIOS folder describe how the AI should interact with his notes in terms any model can understand.
The system outlasts any particular AI tool. That is the point.