AI Compass / Knowledge cards

Knowledge base for agents

the LLM wiki after Karpathy

Checked 17 October 2026 · Michael Engel · Prompting · Automation

In April 2026, Andrej Karpathy described a pattern, the “LLM Wiki”. It has three layers: raw sources that the model reads but never changes; wiki pages that the model writes from the sources and keeps up to date; and a rules file, such as the CLAUDE.md, that defines how material is ingested and maintained. Unlike plain document search, the knowledge is merged once and then carried forward.

What it is for

Domain knowledge such as standards, offer building blocks or notes sits in one place that Claude Code reads and updates. You don’t have to explain it again in every chat. Knowledge about the world stays separate from memory about you and your projects.

Steps

1. Read Karpathy’s text and let Claude Code set up the folder following this pattern. Then check the structure.
2. One knowledge area per wiki.
3. Write the ingestion rule into the folder’s CLAUDE.md: read the new file, discuss it with you, write a source page, update affected pages, flag contradictions. Note the origin for each source page.
4. For larger collections, organize: one topic per file, a short header with name and description, an index.md per folder. State in the CLAUDE.md that the indexes are read first. Without this hint, according to AI LABS, Claude falls back to keyword search.
5. Test queries and check whether the model really searches in the wiki.
6. Every few weeks, have it check for contradictions, outdated statements and orphaned pages. Karpathy calls this “lint”.
7. Only once that works, have material ingested automatically at regular intervals, with write permission only in the wiki folder.

Limits

• Unchecked sources make the wiki worse. What goes in shapes every later answer.
• According to Karpathy, the approach with index files holds about 100 sources and a few hundred pages. For larger, unchanging collections such as transcripts or books, Jack Roberts recommends a vector search, a search by meaning instead of by keyword.
• “Improves itself” is a claim by the creators.
• The video title “Google Replaced The Karpathy’s LLM Wiki” exaggerates. Google’s Open Knowledge Format specification organizes the pattern; it doesn’t replace it.
• Confidential documents only once confidentiality is settled.

Links

• Keeping context clean: reading indexes instead of loading everything saves context.
• Maintaining your Claude Code setup: the wiki’s rules file needs the same upkeep as any CLAUDE.md.
• Web data: settle the access path first: third-party texts that go into the wiki can contain injected instructions.
• Three-tier memory: the counterpart for knowledge about you and your projects.

Sources

• Andrej Karpathy: LLM Wiki, GitHub Gist, April 2026, retrieved 11 Oct 2026
• Jack Roberts: Build a Hermes Knowledge Base That Self-Improves, 14 June 2026, 3:15–6:31, 12:21–13:24. Self-promotion (community)
• AI LABS: Google Replaced The Karpathy’s LLM Wiki, 26 June 2026, 5:20–5:52, 9:05, 10:11. Sponsored by Mobbin
• Jack Roberts: Claude Code + Karpathy’s Obsidian = New Meta, 10 Apr 2026, 8:04–9:08, 16:04–17:11. Self-promotion (community)

Status

Checked 11 Oct 2026 · Maturity new (pattern openly described, not tested by me)

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab