LLM-Maintained Wikis Turn RAG Into Compounding Memory

·https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
repoideapatternaillmragobsidianvaultmemoryagent-loops

Maps to joelclaw Brain/pi-notes as the agent-maintained middle layer between raw sources and reusable answers.

Andrej Karpathy’s LLM Wiki is not another retrieval-augmented generation wrapper. It’s a portable idea file for an agent-maintained Obsidian wiki where raw sources stay immutable and the LLM owns the messy middle: summaries, entity pages, cross-links, contradiction notes, index updates, and logs.

That’s clever because it moves the expensive work from query time to maintenance time. NotebookLM and ChatGPT file uploads can answer from a pile of docs, but Karpathy is pointing at a compounding artifact: every ingest updates the map, every useful query can be filed back into the map, and lint passes keep stale claims and orphan pages from quietly rotting.

This maps cleanly to joelclaw and the Obsidian / agent workflow we keep circling. Not as a giant vector-store religion, but as a small discipline: immutable sources, LLM-written Markdown, a schema file like AGENTS.md or CLAUDE.md, an index.md, and an append-only log.md. It’s memory as a maintained codebase, not chat exhaust.

Key Ideas

  • Raw sources stay immutable; the agent reads them but does not mutate the source of truth.
  • The wiki is generated Markdown that the LLM owns: summaries, entity pages, concept pages, comparisons, cross-links, and synthesis.
  • Schema files such as AGENTS.md and CLAUDE.md turn a generic agent like OpenAI Codex, Claude Code, OpenCode, or Pi into a disciplined maintainer.
  • Ingest is not “store this for later”; it is “update the knowledge graph now,” often touching many pages from one new source.
  • Query results can become durable pages too, so analysis and comparisons do not vanish into chat history.
  • Lint passes are first-class maintenance: contradictions, stale claims, orphan pages, missing cross-references, and missing concept pages get surfaced instead of ignored.
  • index.md and log.md are boring on purpose: one gives content navigation, the other gives chronological receipts that simple Unix tools can parse.
  • Search can stay simple until the wiki grows; qmd adds local hybrid search, LLM re-ranking, a CLI, and a Model Context Protocol server when needed.
  • The lineage is Vannevar Bush’s Memex, especially “As We May Think”: associative trails matter, but somebody has to maintain the damn trails.

Links

← All discoveries