LLM-Wiki 2026: How Karpathy's Patterns Revolutionize the Personal Knowledge Base

Andrej Karpathy, co-founder of OpenAI and former AI director at Tesla, published an unassuming GitHub Gist in April 2026 titled “llm-wiki”. It was neither a product launch nor a scientific publication – merely an idea he had intended as a template for AI agents like Claude Code or Codex. The response was overwhelming: within a few days, the Gist had garnered over 5,000 stars, thousands of forks, and dozens of reimplementations on GitHub. VentureBeat, DataCamp, and numerous tech blogs picked up the concept. What lies behind the hype, and why could the LLM Wiki approach change the way we manage knowledge with AI?

The Problem: Why Knowledge Databases Fail

Every knowledge worker knows the phenomenon: a new Notion project or Obsidian vault starts with enthusiasm – after a few months, hundreds of pages lie fallow, cross-references are outdated, summaries don’t exist. The reason is simple: collecting sources is easy, organizing is hard, and maintaining them over the long term becomes practically impossible as volume grows. Manually updating cross-references, spotting contradictions, and creating summaries costs more time than most are willing to invest.

Karpathy’s observation: LLMs are predestined for exactly this kind of bookkeeping work. They read documents, identify key concepts, create summaries, generate cross-references, and spot contradictions – effortlessly, consistently, and at negligible cost. Humans curate what goes in; the LLM does the rest.

The Three-Layer Architecture

The LLM Wiki pattern defines three clearly separated layers:

1. Raw sources (raw/): A collection of immutable source documents – articles, papers, code repos, data. The LLM reads from this folder but never modifies it. Every claim in the wiki can be traced back to a source file.

2. The Wiki (wiki/): A directory of LLM-generated Markdown files – concept pages, entity pages, source summaries, comparison tables. The LLM owns this layer completely: it creates pages, updates them when new sources are added, maintains cross-references, and detects contradictions. Two special files aid navigation: index.md (table of contents, updated with every ingest) and log.md (appending operation log).

3. The Schema (CLAUDE.md or AGENTS.md): The most important file of the entire system. It defines the wiki structure, naming conventions, page templates, and workflows. It turns a generic LLM into a disciplined knowledge worker.

The Three Operations: Ingest, Query, Lint

The pattern defines three core operations:

Ingest: The user places a new source in the raw/ folder and asks the LLM to process it. The agent reads the document, discusses the key points, writes a summary page, updates the table of contents, and revises all affected concept and entity pages – typically ten to fifteen files per source.

Query: The user asks a question against the wiki. The agent searches the table of contents, reads the relevant pages, and synthesizes an answer with source citations. Important insight: good answers can flow back into the wiki as new pages, further enriching the knowledge base.

Lint: At regular intervals, the agent checks the health of the wiki: contradictions between pages, outdated claims, orphan pages with no incoming links, missing cross-references, and data gaps that could be filled by a web search.

LLM Wiki vs. RAG: A Fundamental Difference

The crucial difference between classic RAG and an LLM Wiki lies in the timing of knowledge work. RAG systems retrieve relevant text fragments from a vector index on every query and synthesize an answer from them – anew each time, with nothing accumulating. The LLM Wiki, by contrast, compiles the knowledge already when a source is added. The synthesis work happens once and persists; the query then reads from an already thought-through foundation.

Both approaches have their place. RAG is better suited for highly dynamic data (e.g., current support tickets) and scenarios where low latency matters more than deep understanding. The LLM Wiki shines with long-lived, accumulating knowledge – research topics, product documentation, strategic analyses – and in practice is often a complement: once the wiki grows beyond a certain size, RAG can be layered on top of the wiki pages themselves.

Karpathy himself reported that his wiki on a single research topic had grown to about 100 articles and 400,000 words – longer than most dissertations – without him having directly written a single sentence.

Conclusion

The LLM Wiki pattern is more than just another AI hype. It addresses a fundamental problem of knowledge work: the maintenance burden that causes almost all personal knowledge databases to fail. By letting LLMs handle the tedious bookkeeping work and assigning humans the role of curator and discoverer, it creates a model for knowledge that grows richer with every new source, rather than starting from scratch on every query.

The open-source community has enthusiastically embraced the concept: within a few weeks, dozens of implementations for Claude Code, OpenAI Codex, and Gemini CLI emerged. Integration with Obsidian as a frontend is a recurring pattern. Anyone looking to take their personal knowledge management to the next level will find in the LLM Wiki a practical, locally runnable approach – beyond cloud dependencies and proprietary RAG platforms.

Sources

• Karpathy, A.: „LLM Wiki" – GitHub Gist (April 2026)
• DataCamp: „LLM Wiki: A New AI Knowledge Architecture in 2026" (Juli 2026)
• Starmorph: „How to Build Karpathy’s LLM Wiki – The Complete Guide" (April 2026)
• Denser.ai: „LLM Wiki: Karpathy’s Idea for AI Knowledge Bases" (April 2026)