An LLM wiki that maintains itself.

On 4 April 2026 Andrej Karpathy published an idea file, LLM Wiki, in three layers: raw sources you never edit, a wiki of pages the model writes and keeps current from them, and a short file of instructions. Three jobs run on it: ingest new sources, answer questions from the wiki, and lint the wiki for contradictions and gaps.

His own setup is an Obsidian vault, with the schema in CLAUDE.md or AGENTS.md, a folder layout and prompts you run by hand. It works, and you maintain all of it. In Morion your notes are the sources, the topic documents Mo keeps are the wiki, and the instructions live where your agents already read them. Ingest and topic upkeep run on their own; contradictions and gaps come back with every answer.

Create a project and a folder for it. Import the Markdown you have: an Obsidian vault folder, meeting notes, specs, a CLAUDE.md, exported pages. Or paste text into new notes; each note is a source.

Keep writing as you work. A note does not need a template; Mo reads what is there.

Open Folder Settings → AI Access and turn on AI access for the folder. Indexing runs on the Claude Code or Codex login you already have, so there is no key to set up: Mo on the subscription you already pay for.

Mo reads every note in the folder, writes a summary and keywords for each, sorts the notes into the project, and groups them into topics. Folders you hide from AI stay out of it.

Each topic has a topic document: what the notes say about the subject, with links back to them. That is the wiki layer. Mo keeps topic documents current as notes change. Your own words outside Mo’s sections stay as you wrote them, so edit a topic’s overview when you want to steer it. Regenerate a topic by hand when you want it now.

Ask Mo through your agent over MCP; that runs on the same login. In the app, the Ask Mo chat needs Pro and a key-based model. Every claim in the answer names the note it came from and carries a status, such as decided, provisional or superseded. Two notes that disagree come back as a conflict; a question the notes do not cover comes back as a gap. That covers the part of the lint step you need when you ask: contradictions and gaps show up in the answer, with the notes behind them.

The window shows the same kind of answer in Mo Assistant, the in-app chat on Pro.

Connect Claude Code or Codex in Settings → Integrations → Apps and Agents. From then on the agent reads the wiki before it works and writes back what it learned: a decision as a note, a finding as a comment on a ticket. Mo folds new notes into the topics. Karpathy files good answers back into the wiki; here the agent writes them as notes.

Set folder permissions so agents see only the folders you allow, and can create notes only where you want them. To keep a folder out of Mo, turn off AI access for it in Folder Settings; that hides it from your agents too.

A wiki that keeps up with your notes without a schema file or a routine: sources in, topics built, answers with the note behind every claim, and contradictions shown instead of repeated. Any agent you connect reads the same wiki and adds to it.

Coming from Obsidian? Morion vs Obsidian.