llm-wikid

An AI-maintained knowledge base that lives in Obsidian. Based on Karpathy's LLM Wiki pattern.

You dump raw sources into a folder. An AI agent reads them, compiles structured wiki pages with cross-references, runs bias checks, and maintains a master index. Every question you ask gets filed back in. The wiki compounds the more you use it.

Works with any agent that can read markdown and run shell commands: Claude Code, OpenClaw, Hermes, Codex, or your own setup.

Quick Start

The agent reads CLAUDE.md and knows everything. That file is the schema that controls the entire system.

How It Works

This is not RAG. RAG re-derives answers every time by chunking documents and running vector search. This system compiles knowledge once into structured wiki pages, keeps them current, and pre-builds cross-references. At ~100 articles / ~400K words, Karpathy found this outperforms RAG for Q&A.

The compound loop: every answer filed back means the next query has more to work with. Your curiosity makes the system smarter.

The Ingest Pipeline

When you run /wiki-ingest, the agent processes everything in raw/ through these phases:

Phase 0 - Sort clippings. Anything in raw/clippings/ gets moved to the right subfolder based on its URL. Tweet goes to raw/bookmarks/, article to raw/articles/, PDF to raw/papers/. The clippings folder is just an inbox.

Phase 1 - Resolve. The agent detects what each file contains and fetches the full content: