LLM Wiki

LLM Wiki is Andrej Karpathy’s pattern for an LLM-maintained markdown knowledge base that compounds between queries instead of re-deriving answers via RAG. We trial it under Technique as the abstract idea behind agent-curated vaults like this one.

Blurb

A pattern for building personal knowledge bases using LLMs.

Summary

When to use: You want knowledge to accumulate across sessions: ingest sources once, maintain entity and concept pages, and query a compiled wiki rather than raw RAG every time.

When to skip: One-shot Q&A over a fixed corpus is enough, or you will not keep sources and a schema (AGENTS.md / CLAUDE.md) under version control.

Core loop:

Three layers: immutable raw sources; LLM-owned wiki markdown; a schema doc that disciplines the agent.

Vs related notes: gbrain is an opinionated implementation of similar ideas. Open Knowledge Format (OKF) is an interchange spec that formalizes markdown knowledge bundles. This note is the pattern itself.

Details

Architecture (From the Gist)

1. Raw sources - curated, immutable inputs the agent reads but never edits.
2. The wiki - LLM-generated markdown (summaries, entities, concepts, synthesis). Humans read; the agent writes.
3. The schema - conventions and workflows in AGENTS.md, CLAUDE.md, or equivalent. Co-evolved with the agent.

Indexing and Logging

• index.md - content catalog (links, one-line summaries, categories). Agent reads it first on query.
• log.md - append-only chronology of ingests, queries, and lint passes.

Why It Differs From RAG

RAG rediscovers fragments on every question. An LLM wiki compiles once and keeps synthesis current: cross-links, contradictions, and topic pages are already maintained. Obsidian (or similar) is the IDE; the LLM is the programmer; the wiki is the codebase.

Optional Tooling Called Out in the Gist

Local markdown search (e.g. qmd), Obsidian Web Clipper, graph view, Dataview over frontmatter, Marp for slides. All optional; the pattern stays file-based.