Andrej Karpathy has released the 'LLM Wiki' pattern, in which you simply drop raw materials into a folder and an LLM automatically builds and maintains an interlinked Markdown knowledge base. Its defining trait is that the user, in principle, never writes the wiki by hand.
April 4, 2026 · Andrej Karpathy
LLM Wiki: Drop Raw Sources In, Get a Self-Maintaining Knowledge Base
A personal knowledge-base pattern where you dump raw articles, papers, repos and images into a folder — and an LLM reads, summarizes, and maintains a cross-linked Markdown wiki. No hand-cleaning, no writing. Pitched as "the next step after RAG."
~60s
to scaffold the setup by pasting a schema
~0
maintenance — knowledge compounds itself
~100
sources / few hundred pages before adding search
What accumulates over time?
Conventional RAG rediscovers from scratch every query. LLM Wiki builds a persistent, compounding layer.
Typical RAG
Nothing persists — starts over each time
syntheses
overviews
entity pages
summaries
LLM Wiki
Persistent wiki layer keeps compounding
The workflow — raw/ stays immutable, wiki/ is LLM-built
Ingest
Bring in sources, update the wiki & links
→
Query
Ask the wiki, write answers back in
→
Lint
Detect contradictions, gaps & staleness
raw/ · immutable sources
wiki/ · [[wiki-links]] pages
CLAUDE.md · schema & style
index.md / log.md · catalog & log
Well received
Scaffolds in ~60s, near-zero upkeep
Great with Obsidian graph view
Ingest/lint as Claude Code commands
Local-first, low-overhead, Git = free versioning
Caveats & criticism
Large first ingest hits tokens & rate limits
Hallucination / structural breakdown risk
Weak cross-device & multi-provider links
Called "an overhyped AI Zettelkasten"
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…