Concept: OBSIDIAN / Karpathy LLM Wiki Pattern
Category: methodology / system architecture Source: Andrej Karpathy GitHub gist (April 2026) — https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f Go Beyond implementation: This wiki (OBSIDIAN)
What It Is
The Karpathy LLM Wiki pattern treats AI knowledge management like code compilation rather than interpretation. Instead of RAG (re-deriving answers from raw docs every query), the LLM builds and maintains a persistent, structured wiki from raw sources.
The compiler analogy:
- Raw sources (articles, transcripts, files) = source code
- The LLM = compiler
- The wiki = compiled executable
Compile once. Query the compiled output. Knowledge compounds over time.
Three-Layer Architecture
| Layer | Description | Who Owns It |
|---|---|---|
| Raw sources | Immutable inputs — transcripts, files, articles | Human (curate only) |
| The wiki | Structured, interlinked markdown files | LLM (creates & maintains) |
| Schema (CLAUDE.md) | Operating instructions and conventions | Human (sets once, refines) |
Three Operations
- Ingest — New source → LLM writes summary, updates 5–15 wiki pages, cross-links
- Query — Question → LLM reads relevant pages, synthesises answer, files good answers back
- Lint — Health check → contradictions, orphans, stale pages, missing cross-refs
Why It Beats RAG for Go Beyond
| RAG | OBSIDIAN |
|---|---|
| Re-derives from raw docs every query | Compiled knowledge, instant synthesis |
| No accumulation | Compounds with every session |
| Retrieval quality depends on chunking | Synthesis quality improves over time |
| Humans must curate and retrieve | LLM curates, maintains, cross-references |
| Good answers disappear into chat | Good answers get filed back permanently |
Go Beyond Implementation Notes
- Installed: 2026-05-01 on machine DAMIEN
- Location: the vault (R2) (accessible across all machines with OneDrive)
- Multi-machine tracking via machine-log.md
- Slash commands: /OBSIDIAN-ingest, /OBSIDIAN-query, /OBSIDIAN-lint
- Grows from: Cowork sessions, uploaded files, external references
See Also
- 2026-05-01-decision-implement-OBSIDIAN — Decision to build this
- company — Go Beyond context