LLM-Wiki 2026: From Gist to Ecosystem – Obsidian Plugin, Enterprise Adoption, and Six Months of Community Practice

It has been six months since Andrej Karpathy published an unassuming gist on GitHub—a few hundred lines of prompt patterns, no installer, no repository. The idea: Instead of searching through documents anew for every question like with RAG, an AI agent builds a permanent, linked knowledge base in Markdown from the sources—the “LLM Wiki.” What seemed like an experiment for AI enthusiasts back then has developed into one of the most discussed patterns in personal knowledge management. With 5,000 stars and just as many forks, it is becoming clear: an ecosystem is emerging here.

From Prompt Pattern to Platform: The Obsidian Plugin

The most influential further development of the Karpathy pattern is an Obsidian plugin that integrates the gist into a full-fledged editor. Developed by gd4ai and Greener-Dalii, the plugin reached 45,000 downloads within five months and is currently available in version 1.27.1. The decisive difference compared to other implementations: It is a pure Obsidian plugin without external dependencies. No Python runtime, no vector embedding model, no separate desktop app—the entire workflow lives within the editor.

The architecture relies on PPR (Personalized PageRank) plus Monte-Carlo search over the [[wiki-link]] graph instead of conventional RAG. The plugin supports over 16 AI providers (Anthropic, OpenAI, Gemini, DeepSeek, Qwen, Ollama, LM Studio, OpenRouter, etc.), and the data never leaves the device when a local model is used. PDFs, Office documents, and images are processed natively. The community has since translated the plugin into eleven languages.

The 200-File Limit and Its Solution

One of the most interesting insights from months of practice comes from Kunal Ganglani, who used the Karpathy pattern daily for three months and documented his experiences in detail. His central finding: Once the wiki exceeds around 150 to 200 files, most AI agents can no longer keep the complete graph in context. The quality of links and updates noticeably suffers.

The community has developed a pragmatic workaround for this: a master index that lists every page with a one-line summary. The agent first reads the index, then selectively loads only those pages relevant to a specific update. In practice, this approach extends capacity to over 300 pages. For even larger knowledge bases, qmd exists as a CLI tool that offers local hybrid search (BM25 plus vector with re-ranking) as an MCP server—essentially RAG over the wiki when the wiki itself becomes too large.

Enterprise Adoption: When the Pattern Meets the Organization

The natural question after six months: Can the personal pattern be transferred to companies? Falconer, a company specializing in knowledge management, has taken on this question and identified the four properties that make Karpathy’s wiki successful: Capture, Link, Compound, and Stay Current.

The analysis shows: None of these properties can be scaled directly to enterprise size. The personal raw/ folder, where the user manually curates sources, has no equivalent in a company. Relevant sources are distributed across GitHub, Slack, Linear, Confluence, Google Drive, and dozens of other tools. Bidirectional links would have to work across tool boundaries—a Slack decision would need to be linked to the implementing PR, the Linear ticket, and the meeting transcript. And the automatic health checks that Karpathy regularly runs on his personal wiki would have to function in an organization without a manual curator.

Y Combinator explicitly named this missing foundation in the Spring 2026 RFS: “A Company Brain that AI agents can actually use”—an infrastructure that does not exist today. The Stack Overflow Developer Survey 2024 demonstrates the urgency: Over 60% of developers spend at least 30 minutes daily searching for solutions, and 68% encounter a knowledge silo at least once a week. Among managers, the figure rises to 73%.

Comparison of Implementations

In six months, the LLM wiki ecosystem has produced several competing approaches: SamurAIGPT/llm-wiki-agent (around 3,300 GitHub stars, as a Claude‑Code/Codex skill) relies on Louvain community detection and SHA256 caching, nashsu/llm_wiki (10,000+ stars) provides a Tauri desktop app, and atomicstrata/llm-wiki-compiler works as a TypeScript CLI with BM25 plus semantic search over chunks. The Obsidian plugin remains the only one that works without an additional runtime.

Conclusion

The LLM Wiki has left the proof of concept behind. The Obsidian plugin with 45,000 downloads proves that the pattern works productively for individual users—with limitations starting at 200 files, which are, however, addressed by community solutions. The more exciting question is enterprise adoption: Here, the infrastructure is still missing to automatically scale the four success properties to enterprise size. Y Combinator has recognized the need—whether and when the first satisfactory solution arrives will be one of the most interesting topics in the coming months.

Sources

• LLM Wiki Setup: Karpathy’s Knowledge Base [2026 Guide] – Kunal Ganglani — Three-month practice report with the 200-file limit and workarounds
• Karpathy LLM Wiki – Obsidian Plugin — Plugin documentation with comparison table and architecture description
• The Enterprise LLM Wiki – Falconer Guides — Analysis of the four properties and their scalability to companies
• Karpathy’s LLM Wiki Gist — Original prompt pattern (5,000+ stars, 5,000+ forks)
• Stack Overflow Developer Survey 2024 — User data on knowledge silos and search times