LLM Wiki 2026: How Karpathy's Pattern Redefines Knowledge Work with AI

In April 2026, Andrej Karpathy, co-founder of OpenAI and former AI Director at Tesla, published a GitHub Gist titled “LLM Wiki” – a concept that went viral within days and now has over 5,000 stars and forks. The core idea is surprisingly simple: instead of extracting knowledge anew from raw documents for every query (as RAG does), an AI agent builds and maintains a persistent, structured knowledge store made of linked Markdown files. Knowledge is compiled once and then continuously updated – a “compounding artifact” that becomes richer with every source and every question.

The Three-Layer Architecture

Karpathy’s LLM Wiki consists of three clearly separated layers:

Raw Sources: An immutable directory of original documents – PDFs, articles, web clippings, notes. The agent reads from it but never changes the source. This is the “Source of Truth” that allows recompilation at any time.

The Wiki: A folder of LLM-generated Markdown files. Each concept, entity, and topic gets its own page. The agent creates and updates these pages, maintains cross-references ([[Wiki Links]]), and identifies contradictions between sources. Karpathy himself reports that his wiki has grown to about 100 articles and 400,000 words without becoming cluttered.

The Schema: A configuration file (e.g., CLAUDE.md for Claude Code or AGENTS.md for Codex) that defines the rules for the agent: How is a new source ingested? What page format should be used? How are references set? The crucial point: the schema is the actual institutional memory – the pages can be regenerated at any time; the schema remains.

The Three Operational Modes

Karpathy describes three key operations that keep the system running:

Ingest: A new source lands in the raw directory. The agent reads it, discusses the core claims with the user, writes a summary page, updates the 10 to 15 existing pages affected by it, and adds an entry to the change log. A single source can thus change dozens of pages in the wiki – the linking work happens once during ingestion, not on every query.

Query: The user asks a question. The agent searches for relevant wiki pages, reads them, and synthesizes an answer with source citations. Importantly: good answers can flow back into the wiki as new pages – comparisons, analyses, or newly discovered connections do not disappear into the chat history.

Lint: At regular intervals, the agent checks the wiki for contradictions, outdated claims, orphan pages with no incoming links, and missing cross-references. The lint run keeps the wiki consistent even with hundreds of pages – a task that in human-maintained wikis is usually neglected after a few weeks.

LLM Wiki vs. RAG: A Paradigm Shift

The fundamental difference lies in the timing of synthesis:

Dimension	Traditional RAG	LLM Wiki
Synthesis timing	On every query	On ingestion
Knowledge state	Stateless, every query starts from zero	Persistent, grows with each source
Cross-references	Implicit (embedding similarity)	Explicit (Markdown links between pages)
Contradictions	Not detected	Actively flagged
Infrastructure	Vector database, embedding model	Only Markdown folder + LLM Agent
Scaling	Millions of documents	Hundreds of pages (context-window limited)

RAG scales horizontally with more documents. The LLM Wiki scales vertically with deeper synthesis. For searching 100,000 support tickets, RAG is the right choice. For synthesizing 50 research papers into a coherent understanding, the LLM Wiki wins.

Tools and Community Implementations

Since Karpathy’s publication in April 2026, the community has produced several implementations:

• llm-wiki (dudarik.com): A tool that turns Claude Code into a wiki manager. It compiles sources into Obsidian-compatible, cross-referenced Markdown files and keeps them up to date.
• nashsu/llm_wiki (GitHub): A cross-platform desktop application that automatically transforms documents into an organized, linked knowledge network.
• LLMWikiNG: A local, privacy-friendly platform that compiles documents once into an OKF (Open Knowledge Format) compliant Markdown wiki.
• ScrapingArt/Karpathy-LLM-Wiki-Stack (GitHub): A comprehensive reference implementation for Obsidian and Claude Code that claims to be ready to use in 15 minutes.

Challenges and Limitations

As convincing as the concept is, there are practical limits. From about 150 to 200 pages, most agents can no longer keep the entire wiki in the context window. The workaround is a master index file that lists each page with a one-line summary. The agent first reads the index and then loads only the needed pages – this extends practical capacity to over 300 pages.

Additionally, the LLM Wiki is not suitable for real-time data or highly dynamic environments. Each new source requires a compilation step – unlike RAG, which makes freshly indexed documents available immediately.

Conclusion

The LLM Wiki is more than another AI trend. It is an architectural realignment: instead of indexing raw data and synthesizing it anew for every query, knowledge is compiled once and then kept alive. The community has translated the concept from an idea into production-ready tools in record time. For anyone who wants to dive deep into a topic, build a research overview, or take their personal knowledge management to the next level – and who is willing to hand over control of the structure to an AI agent – the LLM Wiki is the most promising approach in a long time. The coming months will show whether the pattern also works at enterprise scale or whether the context-window limitations require fundamental further development.

Sources

• Andrej Karpathy: LLM Wiki – GitHub Gist (April 2026)
• DataCamp: LLM Wiki – A New AI Knowledge Architecture in 2026
• Kunal Ganglani: LLM Wiki Setup – Karpathy’s Knowledge Base (2026 Guide)
• AI Builder Club: Karpathy’s LLM Wiki – A Knowledge Base That Compounds
• Tericsoft: LLM Wiki – The Self-Updating AI Knowledge Base