Karpathy's LLM Wiki 2026: How AI Revolutionizes Personal Knowledge Databases

In April 2026, Andrej Karpathy, co-founder of OpenAI and former AI director at Tesla, published a short GitHub Gist titled “LLM Wiki”. The post went viral – over 16 million views on X (Twitter) and more than 5,000 stars on GitHub within days. What Karpathy described there was not a new product, but a workflow pattern: instead of using LLMs just for chatting or code generation, they should build and maintain personal knowledge bases in real time. The idea struck a chord because it solved a problem every knowledge worker knows: knowledge bases that collapse under their own maintenance load.

What is the LLM Wiki?

Karpathy’s core idea is surprisingly simple. Instead of searching through raw documents anew for each query and stitching together answers as in RAG (Retrieval-Augmented Generation), the LLM compiles the knowledge once and keeps it up to date – like a compiler that translates source code into an executable program. The result is a folder of structured, linked Markdown files that an LLM agent automatically creates and maintains.

“Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase,” Karpathy describes the division of roles in his Gist. The human curates sources, asks questions, and thinks about connections. The LLM does the heavy lifting: summarizing, creating cross-references, flagging contradictions, and keeping pages current.

The Three-Layer Architecture

The LLM Wiki consists of exactly three folders:

1. raw/ – The immutable sources. Articles, PDFs, code repos, images, and datasets go here. The LLM reads these files but never modifies them. They serve as the permanent source of truth (Starmorph, 2026).

2. wiki/ – The LLM-generated wiki. This directory contains all compiled Markdown pages: concept pages, entity pages, source summaries, and comparison tables. Two special files are central: index.md – a table of contents of all pages that the LLM updates with each ingestion, and log.md – an append-only log of all operations (Karpathy, 2026).

3. The schema file (CLAUDE.md for Claude Code, AGENTS.md for Codex). It defines naming conventions, page templates, and workflows. It turns a generic LLM into a disciplined knowledge worker (Starmorph, 2026).

The Three Core Operations

Karpathy defines three operations that govern daily life with the LLM Wiki:

Ingest: A new source lands in raw/. The LLM reads it, creates a summary, updates 10 to 15 related wiki pages, creates new concept or entity pages as needed, and logs the process. A single ingest can affect dozens of pages (Kunal Ganglani, 2026).

Query: The user asks questions against the wiki. The LLM navigates via index.md to the relevant pages, synthesizes an answer, and outputs it with [[wiki-link]] citations. Important insights can flow back into the wiki as new pages – knowledge compounds (Karpathy, 2026).

Lint: A regular health check. The LLM looks for contradictions between pages, orphan pages with no incoming links, outdated claims, and knowledge gaps. Starmorph compares this to eslint for knowledge.

LLM Wiki vs. RAG: Two Fundamentally Different Approaches

The confusion between LLM Wiki and RAG is understandable – both work with LLMs and documents. However, the architectural difference is fundamental:

Dimension	Traditional RAG	LLM Wiki
Synthesis time	At each query (query time)	At ingestion (ingest time)
Knowledge state	Stateless – each query starts fresh	Persistent – compounding artifact
Cross-references	Implicit (embedding similarity)	Explicit (Markdown links)
Contradictions	No handling	Active flagging on ingestion
Infrastructure	Vector DB, embedding model, pipeline	Markdown folder + LLM agent
Max. scaling	Millions of documents	Hundreds of pages (context window limited)

The key trade-off: RAG scales horizontally with more documents. The LLM Wiki scales vertically with deeper synthesis (Kunal Ganglani, 2026). For a personal research knowledge base with 50 to 200 sources, the LLM Wiki approach is superior to pure RAG because connections become explicit and navigable.

Community and Tools

The community quickly adopted and extended Karpathy’s idea: The open-source project llm_wiki by nash_su (GitHub) is a full desktop application with a Tauri+Rust backend that implements the pattern as an installable program – including a three-column layout, MCP server integration, and multimodal PDF processing. It already has over 16,600 stars on GitHub (as of August 2026).

Another real-world implementation is “Farzapedia”: Developer Farza fed 2,500 personal diary entries, Apple Notes, and iMessage conversations into an LLM and had it create a 400-article personal encyclopedia with dense cross-references – exclusively as a navigation base for his AI agent (ToolMesh, 2026).

Conclusion

The LLM Wiki is more than just another AI trend – it is a paradigm shift in personal knowledge management. Instead of using LLMs as pure answer machines, they become knowledge compilers that not only find information but condense, link, and make it permanently available. Karpathy’s crucial insight is that the heavy lifting – maintaining cross-references, writing summaries, detecting contradictions – is exactly the task for which LLMs are predestined, while humans can focus on what they do best: curating sources, asking critical questions, and recognizing connections. The LLM Wiki will establish itself alongside established RAG approaches as a valuable tool for anyone who wants not just to manage their knowledge but to truly penetrate it.

Sources

• Karpathy, A. (2026). LLM Wiki – GitHub Gist – Karpathy’s original description of the LLM Wiki pattern
• Kunal Ganglani (2026). LLM Wiki Setup: Karpathy’s Knowledge Base – 2026 Guide – Detailed analysis with long-term experience and comparison to RAG
• Starmorph / Dylan Boudro (2026). How to Build Karpathy’s LLM Wiki – Complete Guide – Complete guide with three-layer architecture and setup
• ToolMesh (2026). Karpathy’s LLM Wiki: Next-Gen Personal Knowledge Management – Overview with Farzapedia case study
• nashsu/llm_wiki – GitHub (2026) – Open-source desktop app with 16,600+ stars implementing Karpathy’s pattern