Concepts Sources Graph Archive Round Robin →

LLM Wiki

Source·published 2025-12 (gist last-modified — original publication date)·GitHub Gist (karpathy)
Open original ↗

Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Summary

Karpathy's gist describes a pattern for personal knowledge bases in which an LLM agent incrementally builds and maintains a structured wiki sitting between the user and the raw source documents. The pattern contrasts itself explicitly with conventional RAG: rather than retrieve fragments from raw sources at query time and re-derive synthesis on every question, the LLM compiles knowledge once into interlinked markdown pages and keeps them current as new sources arrive. The result is "a persistent, compounding artifact" — cross-references already in place, contradictions already flagged, synthesis reflecting everything ingested.

The architecture has three layers: raw/ (immutable inputs the LLM only reads), wiki/ (markdown pages the LLM owns and updates), and a schema file (e.g. CLAUDE.md for Claude Code or AGENTS.md for Codex) that tells the LLM how the wiki is structured and what conventions to follow. The schema is what makes the LLM "a disciplined wiki maintainer rather than a generic chatbot," and the human and LLM co-evolve it as they figure out what works.

Three operations run against the wiki: ingest (read source → discuss takeaways → write/update pages → update index + log), query (read index to scope → drill into pages → return cited synthesis, with the option to file good answers back as new pages), and lint (health-check for contradictions, orphans, stale claims, missing concepts, missing cross-references). Two navigational primitives — index.md (content-oriented catalog) and log.md (append-only chronological journal) — replace the need for embedding-based RAG infrastructure at moderate scale (~100 sources, hundreds of pages).

Karpathy frames the labor division explicitly: the human curates sources, asks questions, and thinks; the LLM does the bookkeeping (cross-references, summary refresh, consistency maintenance) that has historically killed human-maintained wikis. The cost of maintenance approaches zero, so the wiki keeps getting richer instead of decaying. He places this in the lineage of Vannevar Bush's 1945 Memex — "a personal, curated knowledge store with associative trails between documents" — and notes the part Bush couldn't solve was who does the maintenance.

The gist is intentionally abstract: it communicates the pattern, leaves implementation specifics (directory layout, schema conventions, page formats, tooling) to the user and their LLM agent to instantiate together for the user's domain.

Key claims

  1. The wiki is a persistent, compounding artifact, not a query-time retrieval cache. — > "the wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've read."
  2. Three-layer architecture: raw / wiki / schema. — > "There are three layers: Raw sources… The wiki… The schema."
  3. The schema makes the LLM disciplined. — > "This is the key configuration file — it's what makes the LLM a disciplined wiki maintainer rather than a generic chatbot."
  4. Three operations: ingest, query, lint. — > "Ingest. You drop a new source into the raw collection… Query. You ask questions against the wiki… Lint. Periodically, ask the LLM to health-check the wiki."
  5. At moderate scale, index + log replace embedding-based RAG. — > "This works surprisingly well at moderate scale (~100 sources, ~hundreds of pages) and avoids the need for embedding-based RAG infrastructure."
  6. Maintenance is the work, and the LLM does it for free. — > "Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass."
  7. Good query answers can be filed back as new wiki pages. — > "good answers can be filed back into the wiki as new pages."

Concepts cited

Vendors discussed

(none — the gist is pattern-level, not domain-specific)

Sectors touched

(none — pre-domain)

Notable quotes

"Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase."

"The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping."

"The idea is related in spirit to Vannevar Bush's Memex (1945) — a personal, curated knowledge store with associative trails between documents. … The part he couldn't solve was who does the maintenance. The LLM handles that."

"The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else."

Open questions raised

  • The pattern explicitly leaves "implementation specifics … to depend on your domain." For digital-asset vendor research, what does that translate to concretely — entity types, frontmatter shapes, slug rules? (Answered structurally by this vault's CLAUDE.md schema; lived-experience answer accumulates with each ingest.)
  • At what corpus size does the index-only navigation approach break down and require a search tool like qmd?
  • The gist mentions filing good query answers back into the wiki — under what conditions does that warrant a new entity page vs. an append to an existing one?

External references

  • Tolkien Gateway (https://tolkiengateway.net/wiki/Main_Page) — cited as an example of a community-built fan wiki demonstrating depth via interlinks.
  • qmd (https://github.com/tobi/qmd) — local search engine for markdown files; cited as an optional CLI/MCP tool for larger wikis.
  • Vannevar Bush, Memex (1945) — cited as the conceptual ancestor.

Cross-references

  • Andrej Karpathy — gist author.
  • (no other cross-references — this was the first source in the vault; subsequent sources cross-reference the gist as the meta-pattern source.)