The pod vault is Karpathy's LLM wiki pattern: raw sources go in, an agent compiles them into cross-linked pages, the pages get smarter over time, and CLAUDE.md holds the schema. You have been living in one since April. This page is only the two things that turn it from your vault into the crew's.

## Search by meaning

Agents find things by filename and grep. That works until the vault has a thousand notes and the answer is in one titled something you would never guess. The fix is an embeddings index over the markdown and a search endpoint the agent can call.

The small version is an afternoon. Ollama with an embedding model such as `nomic-embed-text`, a script that walks the vault and stores a vector per chunk in SQLite, and a tiny HTTP service that takes a question and returns the top ten chunks with their file paths. No note text leaves the mini. Then one line in CLAUDE.md: "before answering a question about the vault, query the search endpoint first." Jeff's crew runs this shape over several vaults and it is the difference between an agent that knows where things are and one that rediscovers them every session.

If you would rather not build it, the Obsidian plugins that do semantic search work for you in the app; they are less useful for agents, which do not have the app open.

## The rule about what goes in

Nothing secret, ever. The vault syncs, backs up, gets grepped, and gets read into transcripts. A token pasted into a note once is in five places by morning. The crew rule is to redact at the write boundary: anything that writes into the vault from raw material, a transcript miner, a capture pipeline, an agent summarising an email, strips credentials before the write, including from filenames. Cleaning up afterwards never gets all the copies.

The second rule is that the agent commits. End of every task, one commit, message says what changed and why. The git history is the audit trail and the undo.

## How you know it worked

Ask the agent a question whose answer is in a note it has never opened, phrased in words that do not appear in the note's title. It finds it.

## Read more

- [Karpathy's LLM wiki gist](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f)
- [Ollama](https://ollama.com/) and [nomic-embed-text](https://ollama.com/library/nomic-embed-text)
- [Obsidian](https://obsidian.md/)
- [sqlite-vec](https://github.com/asg017/sqlite-vec), if you want vectors in SQLite without a second database