The Karpathy LLM Wiki Pattern

This wiki is an example of the pattern it describes: an AI reads raw material (here, WhatsApp chats), writes and updates a set of linked Markdown pages, and keeps them current as new material comes in. The same idea now sits inside most personal agents as their memory.

The pattern

Andrej Karpathy, the AI researcher, posted it as a short write-up called "LLM Wiki" on 2026-04-04 (the gist). Instead of uploading documents and having the AI search them from scratch on every question, you let it build and maintain a structured wiki.

It has three layers:

• Raw sources that the AI reads and never changes.
• The wiki: pages the AI writes and keeps linked to each other.
• The schema: an instruction file (such as CLAUDE.md or AGENTS.md) that tells the AI how the wiki is organized and how to work on it.

And three jobs:

• Ingest: read a new source and update every page it affects.
• Query: answer from the wiki, and file good answers back as new pages.
• Lint: look for contradictions, stale claims, orphan pages and gaps.

Two files hold it together: index.md, a catalog the AI reads first, and log.md, a running record of what was ingested and checked.

Karpathy's point is that people abandon wikis because the upkeep grows faster than the value. An AI does not get bored of updating cross-references.

How this wiki uses it

The ChabadLabs wiki started from one WhatsApp export of about 7,400 messages over ten months. The October 2026 refresh drew on four shluchim AI groups from October 2025 to October 2026.

That refresh taught the lesson that matters most for anyone copying the pattern: a lot of the "facts" in the source chats were written by members' own AI agents, and many were wrong (invented commands, wrong prices, models that did not exist). So every product fact on these pages is now checked against the vendor's own page before it goes in, and anything that could not be checked is dropped or labeled as one member's experience with a date. An AI-maintained wiki fed from chat needs that rule, or it launders guesses into facts.

The builders' own community wiki (sign-in only) hit the same problem. Its wiki agent refused in August 2026 to publish an operator's pasted claim as general fact "on the strength of a pasted message alone." The same month a corrected proposal and the wrong original were both approved by mistake; the fix was to cancel the old proposal whenever a new one replaces it.

How agents remember things, 2026

The shluchim agent builders tried several approaches. They differ mostly in cost and in who decides what gets saved.

Plain memory files. NanoClaw gives each agent a folder of Markdown notes with an index.md map, loaded into every new conversation. There is no database; you can open and edit the files yourself, and memory stays within one agent group (NanoClaw memory docs). Hermes Agent keeps similar notes and also searches its own past conversations. This is the cheapest option and the right start for a one-person assistant.

A Karpathy wiki skill. GabAI ships a skill that adds this pattern to an agent. Members used one shared wiki for all their agents, or one per area (personal, family, the Chabad House). By design it saves only when you tell it to; ordinary memory handles the rest. The cost warning came quickly: a shliach building a wiki of Chabad texts said it burns tokens "like there's no tomorrow," because each ingest rewrites 10 to 15 pages, and another's agent "wrote 6k wiki pages for the wrong project" in one week (May 2026). The group's advice in June 2026: start with plain memory and conversation logs, and add the wiki only if after four to six weeks the agent still cannot recall what it needs. The full wiki suits a team or community knowledge base.

MemPalace. An open-source (MIT) memory system that stores conversations and notes word for word on your own machine, organized into wings, rooms, closets and drawers, with local search and no API key needed (MemPalace, GitHub). The GabAI admin's view: "Karpathy wiki is good for dealing with large bodies of knowledge. For solid agent contextual long-term memory, MemPalace is better (and cheaper)." He uses it behind the group's support agent and feeds it his Claude Code session transcripts.

Search on top. qmd is a free, local search engine for Markdown notes that combines keyword and meaning-based search and connects to Claude through MCP (GitHub). The admin runs it over his agent's wiki.

Two cautions from the group:

• Wikis built from group chats hold personal information. In May 2026 group admins noticed that an agent could "know" a lot about members from a shared wiki, and one agent mis-answered a question about a member. Ask before you ingest other people's messages, and strip names and numbers you do not need.
• Memory is not a rulebook. An agent can edit its own memory files, so do not put hard limits there. Enforce them in permissions (see Agent Safety).

Where the pattern shows up in Torah apps

• The Maamorim app: a curated corpus with AI presenting it. See Maamorim App.
• dach.dev: plain-text seforim you paste into your AI. See dach.dev.
• NotebookLM with the Rebbe's Torah: source-grounded answers over a fixed set of texts. See NotebookLM.
• A shliach's Ayin Beis site applies Karpathy's wiki to that year's maamarim (May 2026).
• The group's draft principles for AI Torah citation (see Torah Accuracy Debate) are in effect a schema for a Torah wiki.

Limits

• A wiki is only as fresh as its sources. Fast-changing tool facts go stale underneath it. This page's own previous version said nothing about agent memory, and the GabAI facts in the old pages were out of date within weeks.
• It needs regular lint passes, or contradictions and dead links pile up.
• It does not replace the sources. For Torah content, the sefer is the authority; the wiki only points to it.

Related

• AI and Torah Accuracy
• Torah Accuracy Debate
• GabAI, NanoClaw, Hermes Agent
• Agent Safety
• Community apps built on the pattern are in the Toolbox