Karpathy’s llm-wiki
Most work with LLMs and documents is RAG (retrieval-augmented generation): upload files, retrieve matching chunks at query time, generate an answer. It works, but the model rediscovers the same knowledge from scratch on every question. Nothing accumulates.
Andrej Karpathy’s llm-wiki pattern flips that. Instead of retrieving from raw documents at query time, the LLM compiles them once into a persistent wiki, a structured, interlinked set of markdown pages, and reads that compiled layer at query time. The shape most write-ups lead with is three layers: an immutable raw/ folder of your source documents, a wiki/ folder of the LLM-written pages, and a schema file (a CLAUDE.md or similar) that tells the model the conventions:
my-project/
├── raw/ # immutable source documents (the LLM reads, never edits)
│ ├── coscientist-2023.pdf
│ └── benchmark.csv
├── wiki/ # the LLM-written pages: summaries, entity pages, an index, a log
│ ├── index.md
│ └── Coscientist-2023.md
└── CLAUDE.md # the schema: the conventions the LLM follows
Add a source and the model reads it, extracts what matters, and integrates it, updating entity pages, revising summaries, flagging contradictions. Knowledge is compiled once and kept current, not re-derived per query.
The key property is that the wiki is a compounding artifact. The cross-references are already there, the contradictions already flagged, the synthesis already reflects everything read. You never write it; the LLM does the bookkeeping (summarizing, cross-referencing, filing) that makes a knowledge base useful over time. You curate sources and ask questions; the LLM does the grunt work. Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase.
The lineage runs back to Vannevar Bush’s Memex (1945): a private, curated knowledge store with associative trails between documents, an idea that later inspired hypertext and the web. Bush could not solve who does the maintenance: humans abandon wikis because the upkeep grows faster than the value. The LLM does not get bored, does not forget a cross-reference, and can touch fifteen files in one pass, so the wiki stays maintained because maintenance costs almost nothing.
One honest limit: this compile-once approach wins for personal and small-team knowledge bases (hundreds of documents, one curator). At true enterprise scale, millions of files, retrieval-augmented generation is still the practical choice. The wiki pattern is not a universal replacement for RAG; it is the better fit when someone is curating a bounded, evolving body of work. Everything else in these nuggets is what happens when you take it seriously for a research group.
Sources
- Andrej Karpathy, “llm-wiki” (gist), the pattern.
- Vannevar Bush, “As We May Think” / Memex (Wikipedia), the 1945 lineage.