second-brain

llm-wiki-pattern

2 min read

LLM Wiki Pattern

A persistent, LLM-maintained knowledge base where the LLM writes all pages and the human curates sources.

Last updated: 2026-04-12

Overview

Proposed by Andrej Karpathy, the LLM Wiki pattern inverts the typical RAG workflow. Instead of the LLM rediscovering knowledge from raw sources on every query, it maintains a structured wiki of synthesized pages that compound over time.

The core insight: a wiki is a persistent, compounding artifact. Every source added makes the whole richer. Every question asked can leave behind a new page. The human never (or rarely) writes wiki content — the LLM owns that layer entirely.

The human’s job is to: source interesting material, ask good questions, and review the graph. The LLM’s job is to: ingest, summarize, cross-reference, flag contradictions, and keep the index honest.

Key Points

  • Three layers: raw sources (immutable), wiki pages (LLM-maintained), schema/CLAUDE.md (governs behavior)
  • Four operations: INGEST, QUERY, LINT, and ad-hoc NOTE
  • Cross-references are the value — a wiki page with no backlinks is an orphan; the graph view in Obsidian reveals the knowledge structure
  • LLM writes, human curates — reversing typical note-taking labor
  • Schema is config — CLAUDE.md is the “source code” of how the wiki behaves; editing it changes LLM behavior
  • index.md + log.md are infrastructure: the catalog and the audit trail
  • Echoes Vannevar Bush’s Memex concept but solves the maintenance problem Bush couldn’t — LLMs don’t forget to update links

Connections

  • obsidian — recommended editor; graph view, backlinks, and plugin ecosystem make it ideal
  • andrej-karpathy — originator of this pattern

Sources

Graph View

Backlinks