LLMMaintainedWiki
title: LLM-Maintained Wiki type: technique created: 2026-05-19 last_updated: 2026-05-19 related: ["Playradar", "radar/techniques/PersonalWikiSkill", "radar/platforms/ModelContextProtocol"] sources: ["https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
An architecture pattern for building persistent, LLM-curated knowledge bases. Proposed by Andrej Karpathy. The core shift is from retrieval ("retrieve documents on every query") to compilation ("compile knowledge once, keep it current"). The LLM reads source documents, extracts key information, and integrates it into a structured wiki — updating entity pages, revising topic summaries, and maintaining cross-references — rather than re-synthesising raw documents on each query.
| Layer | Content | Mutability |
|---|---|---|
| Raw sources | Documents, articles, PDFs, notes | Immutable |
| The wiki | Markdown files: entity pages, topic summaries, cross-references | LLM-maintained |
| Schema / config | Instructions defining structure and maintenance workflows | Human-maintained |
The wiki layer is the compiled output of repeated LLM passes over raw sources. It accumulates understanding over time rather than regenerating it per query.
Ingest: Process new sources. Update 10-15 wiki pages in one pass. Extract entities, update topic summaries, add cross-references.
Query: Answer questions against the wiki (not raw sources). File valuable discoveries back as new pages during query sessions.
Lint: Periodic health checks for contradictions, orphaned pages, and stale claims.
Index.md provides article navigation. Log.md provides a chronological record of ingested sources. Optional local semantic search via tools like qmd or Obsidian for visualisation.
Traditional RAG retrieves raw documents on every query and synthesises from scratch. LLM-Maintained Wiki pre-synthesises: the wiki pages are the knowledge store, not the raw documents. Query latency is lower; knowledge accumulates and improves across sessions; contradictions and gaps are surfaced during lint rather than during user queries.
LLM-Maintained Wiki sits in the Assess ring of the Techniques quadrant, at inner position. First studied via Karpathy's GitHub gist on 2026-05-19. No direct personal implementation of this architecture. Inner position reflects hands-on exposure through the farzaa Personal Wiki Skill (an active implementation of this pattern in this repo), low friction to trial for any knowledge-intensive project, and the pattern's direct applicability to replacing RAG pipelines where compilation latency is acceptable. The remaining gate is a standalone implementation against a personal or project document corpus.
