What Is Karpathy's LLM Wiki?
Karpathy's compounding knowledge base for AI agents
In April 2026, Andrej Karpathy published a GitHub Gist called llm-wiki.md. Within days it had spread across developer Twitter, Hacker News, and every AI newsletter worth reading.
The idea is simple enough to explain in a sentence. Deep enough that people are still figuring out the implications.
What is an LLM Wiki?
A personal knowledge base made of plain Markdown files, structured so an AI can read, write, and maintain it - not you.
The key shift from normal note-taking: most knowledge bases are built for humans to browse. An LLM Wiki is built for a model to query. That changes everything about how information is organized - flat structure, interlinked entity pages, no folder hierarchies.
Karpathy's framing draws on a basic observation: modern LLMs have context windows large enough to hold an entire personal knowledge base. No vector databases. No RAG pipelines. No embeddings. Just text files and a long context window.
His own framing: "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." You rarely write the wiki yourself. You curate sources, ask questions, think. The LLM handles summarizing, cross-referencing, filing, and bookkeeping.
How does it work?
Three components:
Entity pages - each file covers one concept, person, project, or topic. Structured like a Wikipedia entry: definition, context, relationships to other entities, open questions. Written and maintained by the AI.
Wiki-links - pages reference each other using [[double brackets]]. The AI maintains these cross-references automatically. A page on "context engineering" links to "personal context," "MCP," "RAG" - without you managing it.
Incremental updates - when you add new source material (a paper, a transcript, a set of notes), the AI updates existing pages and creates new ones. Knowledge compounds over time instead of being rediscovered from scratch on every query.
Practical workflow: drop a source into the system, ask the AI to update the wiki, get structured interlinked pages you can query later in plain language.
How is it different from RAG?
RAG embeds your documents as vectors, stores them in a database, retrieves relevant chunks at query time. Useful for large enterprise knowledge bases - tens of thousands of documents where you can't fit everything in context. For personal knowledge bases, it's significant overhead for unclear benefit.
LLM Wiki trades retrieval complexity for pre-compiled structure. Instead of finding relevant fragments at query time, the AI has already organized and cross-referenced everything into readable pages. For personal use - hundreds of pages, not hundreds of thousands - this is simpler and often more useful.
The tradeoff: because the LLM summarizes sources into wiki pages, there's a risk of hallucinations getting baked in as facts. With pure RAG, a wrong answer is just one wrong answer. With an LLM Wiki, a small misunderstanding can propagate quietly across linked pages. Periodic audits matter.
What is it good for?
Topical knowledge - things you want to understand and build expertise in over time:
- Research areas you're tracking (AI safety, distributed systems, a new market)
- Domain knowledge for a project (architecture decisions, technical standards, competitive landscape)
- Any field where you're reading a lot and want the AI to synthesize and connect ideas
The pattern is about building structured understanding of topics - not about telling the AI who you are.
Where does personal context fit in?
LLM Wiki and personal context solve adjacent but different problems.
LLM Wiki answers: what do I know about this topic? A knowledge base about the world - the subjects you're studying, the domains you work in, the ideas you're tracking.
Personal context answers: who am I and what am I working on? A structured profile - your identity, professional background, current priorities, working style - that tells AI tools who they're working with before the conversation starts.
The distinction matters in practice. An LLM Wiki full of architecture notes doesn't help Claude understand that you're a technical founder raising a seed round who prefers direct communication and hates jargon. That's personal context - and it comes from a different source.
Personal context tells every tool who you are. An LLM Wiki tells it what you know. Together, they give an AI tool the foundation it needs to be genuinely useful - not just technically correct.
→ What personal context is and how it works: What Is Personal Context for AI?
→ How to build it from your actual sources: How to Build Personal Context for AI
