other-brain

A personal knowledge vault that compounds as you feed it. You ingest raw sources; an LLM builds and maintains a structured, cross-linked wiki on top of them. You read the wiki, query the wiki, and guide the LLM as it grows.

Inspired by Andrej Karpathy's LLM Wiki pattern. This repo is one practical scaffolding of that idea. Fork it, rename what you want, delete what you don't need.

This is a template, not a product. Everything here is subjective and intentionally incomplete. The whole point is that you edit CLAUDE.md and the folder conventions until they fit how you think.

The problem this solves

The default way to use an LLM against your own documents is RAG. You drop files into a folder, ask a question, and the model chunks, embeds, retrieves, and synthesises on the fly. It works, but it has one big flaw: the model rediscovers your knowledge from scratch on every query. Nothing accumulates.

If a subtle question needs five sources stitched together, that stitching happens again, from zero, on every ask. Your "knowledge base" is really a pile of chunks, not knowledge.

How this pattern fixes it

Two layers, kept strictly separate:

• **raw/ is immutable.** Every source you ingest (article, transcript, PDF, documentation) lives here in its original form. You never edit files in raw/. If a source gets updated, you add a new dated version alongside the old one.