An LLM Wiki is a personal, compounding knowledge base written as plain markdown files that a large language model maintains for you. You drop raw sources — PDFs, notes, screenshots, papers — into a folder, and the LLM compiles them into interlinked wiki pages following a schema you define. Every new source ripples through the existing pages rather than disappearing into a chat log. The pattern was popularized by Andrej Karpathy in April 2026, and it has already displaced parts of how many people use RAG. If you have been trying to figure out what all the fuss is about, this guide explains the LLM Wiki pattern in plain English — what it is, how it differs from traditional retrieval, and how to start your own in the next thirty minutes.
If you came here looking for a tooling comparison, jump straight to our LLM Wiki vs RAG breakdown. Otherwise, read on.
The problem LLM Wiki solves
Most people use LLMs like a vending machine. Ask a question, get a chunk of text, close the chat. The answer goes into a Slack message or a doc, and three weeks later you are asking the same question again. The LLM never remembers; you never accumulate.
RAG (retrieval-augmented generation) tried to fix this by letting the model look things up in a vector database at query time. It works, but it has three stubborn problems:
- Every answer re-derives the same knowledge. The wheel is reinvented every time you ask.
- The output disappears into chat history. You cannot read the wiki — you can only query it.
- Quality is a function of chunking. When chunks are bad, the answer is bad, and you rarely know why.
Karpathy's observation was simple: if you are spending most of your tokens summarizing the same sources into the same answers, maybe you should ask the LLM to write the answers down once and keep them. That is the LLM Wiki pattern in one sentence.
How an LLM Wiki actually works
At its simplest, an LLM Wiki is three folders and one config file:
my-wiki/
├── raw/ # the stuff you dump in
├── wiki/ # the stuff the LLM writes back
├── index.md # the entry point
└── schema.md # your shape of knowledgeYou put raw inputs — PDFs, meeting notes, a YouTube transcript, a paper — into raw/. You point Claude or Gemini at the folder along with your schema.md, and the model reads the new source, decides which wiki pages should be updated or created, and writes them into wiki/. Each page is a plain markdown file with frontmatter and wikilinks, so you can open the vault in Obsidian, grep it in a terminal, or diff it in git.
The schema is the important part. A schema might say "every page is either a concept, a person, a paper, or a tool; concepts link to papers that cite them; papers link to the methods they use." That rule set is what keeps the wiki consistent over hundreds of compilations.
When a new source contradicts an existing page, a well-written schema will tell the LLM to flag the contradiction in a ## Open Questions section rather than silently overwrite. This is where LLM Wiki earns its name — it behaves more like a collaborative encyclopedia than a chat log.
What LLM Wiki is not
Three quick disclaimers before you get excited:
- It is not a SaaS. There is nothing to sign up for. The pattern is a workflow you run against files on your disk.
- It is not RAG, and it is not a RAG replacement for every case. For knowledge bases under roughly 500K tokens, an LLM Wiki is usually simpler, faster, and cheaper. Beyond that, RAG or a hybrid earns its complexity.
- It is not new magic. It is a disciplined way of asking an LLM to maintain a folder of markdown files. The discipline is what makes it work.
Who benefits from an LLM Wiki?
Three groups of people tell us they get the most value:
Independent developers and AI engineers
If you live in Claude Code or Cursor, you probably already have a CLAUDE.md in your projects. An LLM Wiki extends the same pattern to a second brain that covers multiple projects, libraries, and decisions. Your model gets better context every week without you copy-pasting the same design docs into chat windows.
Researchers and grad students
Managing 200+ papers is a different problem from managing 20 notes. A schema-driven LLM Wiki handles the citation graph, the methodology comparison, and the literature-review draft as first-class citizens. If you are drowning in Zotero and half-written Obsidian notes, LLM Wiki for research is the detailed walkthrough.
Writers and no-terminal knowledge workers
You do not need to be technical to benefit. If you already live in Obsidian or Notion and the thought of running a Python script gives you hives, the no-code LLM Wiki guide shows how to run the full pattern through ChatGPT or Claude's web UI. No terminal, no scripts — just drag and drop.
How to start your own LLM Wiki in 30 minutes
Here is the smallest version of the workflow that actually works:
Step 1 — Create the folders. Make raw/ and wiki/ somewhere on your disk. If you use Obsidian, make them inside your vault so you get wikilinks and backlinks for free.
Step 2 — Write a one-page schema. In schema.md, define three to five entity types and the relationships between them. Keep it short — anything longer than a page and you will never iterate on it. If you need a starting shape, we maintain five tested templates for general, research, engineering, product, and SEO use cases on our templates page.
Step 3 — Drop one source in and compile. Put a single PDF, one meeting note, one paper, whatever, into raw/. Open Claude or Gemini and say something like: "Read schema.md, then read raw/the-file.pdf. Update or create the relevant pages in wiki/ following the schema. Do not touch anything else." The first compilation is where you will notice what your schema is missing — that is normal.
Step 4 — Read the output. Open wiki/ and read the generated pages as if you had never seen the source. If anything feels wrong, edit the schema, not the output. Good schemas teach the LLM to write correctly; bad schemas produce plausible-sounding garbage.
Step 5 — Repeat. Add a second source, then a third. Every few weeks, re-read the schema and prune it. Your wiki is healthy when a new source ripples through five or ten existing pages rather than creating an orphan.
That is it. That is the entire LLM Wiki pattern.
Why the LLM Wiki pattern caught on so fast
Karpathy's original gist went viral because it named something a lot of people had been circling for months. Everyone who used Claude Code or Cursor for serious work had discovered that a well-maintained CLAUDE.md made the difference between magic and mediocrity. The LLM Wiki pattern is the generalization of that insight to any knowledge domain, not just software.
The other reason it caught on: it is cheap. Running an LLM Wiki costs you the price of the LLM calls and nothing else. No vector database, no infrastructure, no SaaS subscription. You get a persistent, compounding knowledge base for the price of a few Claude API calls per week. For a personal knowledge worker, that is an absurdly good deal.
Common objections, briefly answered
"Won't the wiki get too big to fit in context?" — For most personal use cases, no. Claude 3.7 and Gemini 2.0 both handle 200K+ tokens, and a disciplined wiki stays well under that for years. When you cross the threshold, a hybrid approach (wiki for structure, RAG for bulk) is the answer. See our LLM Wiki vs RAG comparison for the decision framework.
"Doesn't this lock me into one LLM vendor?" — No. The wiki is plain markdown. The compilation loop is a prompt. Swap Claude for Gemini tomorrow and your wiki is fine; swap it for a local Llama next year and your wiki is still fine. That is the whole point of the markdown-first design.
"Isn't this just note-taking with extra steps?" — It is note-taking where the notes write themselves and stay consistent with each other. If you already have the patience to maintain a hand-written Obsidian vault, you do not need an LLM Wiki. Everyone else does.
Where to go next
If you want the hands-on version, pick the track that matches who you are:
- Developer running Claude Code or Cursor? The LLM Wiki for Developers hub collects every engineering-focused guide, schema, and the Starter Kit.
- Researcher managing 50+ papers? The LLM Wiki for Researchers hub covers Zotero workflows, citation graphs, and the academic schema.
- Writer or knowledge worker who never wants to see a terminal? The LLM Wiki for Writers hub gathers the no-code paths through Obsidian, Notion, and browser-only setups.
Or jump straight to the deep-dive guide that matches your setup:
- Using Obsidian? The LLM Wiki Obsidian setup guide walks you through the vault structure, plugins, and schema you need.
- No terminal, no Obsidian? The no-code LLM Wiki tutorial runs the same pattern through ChatGPT's web UI.
- Managing research papers? LLM Wiki for research covers Zotero integration, citation graphs, and literature-review workflows.
- Curious about the technical debate? Read LLM Wiki vs RAG for when each approach wins.
And if you would rather skip the schema-writing weekend, our curated LLM Wiki templates ship with the exact shapes we use internally — five scenarios, ready to paste into your project.
One more thing: we send an occasional email summarizing new LLM Wiki implementations, schema patterns worth stealing, and the best community discussions. No fixed schedule, no spam, no upsell. Subscribe at the bottom of this page.