Karpathy LLM Wiki Pattern

A knowledge-base architecture popularised by Andrej Karpathy in April 2026 as a RAG alternative for small-to-medium personal corpora. Uses flat markdown files inside a raw/ folder for sources and a wiki/ folder for synthesised entity/concept/analysis pages, plus an index.md and log.md. Claude Code reads, indexes, compacts, and links pages by following markdown links rather than using vector similarity. No embeddings, no vector DB, no chunking pipeline — just markdown and a claude.md that explains the routing. Obsidian is the recommended viewer for the graph view, though not required.

Key points

• Flat markdown structure: raw/ (sources), wiki/ (synthesised pages), index.md, log.md
• Optional subfolders: entities, concepts, sources, analysis, people, organisations
• Hot cache (hot.md) stores recent context summary, useful for executive-assistant style wikis
• Claude Code performs data ingest, organisation, linting, and Q&A phases
• One user dropped token usage 95% vs vector RAG on 383 files + 100 meeting transcripts
• Relationships via markdown backlinks, not similarity search
• Linting phase runs LLM health checks for inconsistent data and gap-filling
• Scales cleanly up to hundreds of pages; vector RAG wins above millions of docs
• Obsidian Web Clipper extension can ingest web articles directly into the raw folder
• In Karpathy's Sequoia interview, the pattern is reframed as an understanding tool: LLMs can project fixed information into new wiki structures, summaries, links, and Q&A surfaces that help a human understand the material better ^[src-055]
• Roberts contrasts the wiki/Obsidian style with Pinecone: use markdown when you want readable, editable files and graph backlinks; use vector memory when scalable semantic search matters more ^[src-059].
• Google Cloud's Open Knowledge Format is an enterprise-facing validation of the same core shape: markdown files, YAML frontmatter, human readability, and agreed conventions so agents can consume knowledge without translation ^[src-100].
• Herk's five-level second-brain framing places the LLM wiki at level two: enough structure to route topics, transcripts, meeting notes, and concept pages, without needing semantic search or graph infrastructure by default ^[src-103].
• The same source separates visual graph browsing from operational usefulness. Obsidian can help humans inspect relationships, but the core requirement is that the agent can read the files, follow the index, and retrieve the right context ^[src-103].

Related entities

• Andrej Karpathy
• Obsidian
• Claude Code
• Jack Roberts
• Pinecone

Related concepts

• Claude Code Context Management Discipline
• Claude Code Token Economics
• Understanding Bottleneck
• Software 3.0
• AI Memory Operating System
• Expert Knowledge Index
• Open Knowledge Format
• Machine-Scannable Content
• AI Second Brain Maturity Model
• LLM Wiki vs Semantic RAG

Source references

• ^[src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)

– Videos referenced: sboNwYmH3AY

• ^[src-013] Nate Herk — "Build & Sell Claude Code Operating Systems (2+ Hour Course)" (2026-05-01)

– Live demo: Karpathy's original prompt used verbatim to build a wiki from the AI2027 article, producing 23 pages (6 people, 5 organisations, multiple concept pages) in one session. Hot.md demonstrated as a session cache. Obsidian Web Clipper shown as the one-click ingestion path for web articles.

• ^[src-055] Sequoia Capital — "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (2026-04-29)
• ^[src-059] Jack Roberts — "This Memory System just 10x'd Claude Code" (2026-05-03)
• ^[src-100] Sam McVeety and Amir Hormati – "How the Open Knowledge Format can improve data sharing" (2026-06-12)
• ^[src-103] Nate Herk – "Every Level of a Claude Second Brain Explained" (2026-06-17)