The LLM Wiki,
your second
brain.
Karpathy shared a pattern in 2026 (a reusable design blueprint) that's simple but powerful: instead of a classic RAG (Retrieval Augmented Generation, in plain English: the AI goes and fetches the answer from your documents on every question), the AI builds and maintains a real wiki that grows richer with every new source. A true second brain that compiles and sharpens itself. Let me walk you through the concept and how to set it up.
What you'll learn
- Why RAG isn't enough — the AI "rediscovers" your knowledge on every query, nothing accumulates
- The LLM Wiki pattern — 3 layers (raw sources / wiki / schema = structure rules), 3 operations (ingest = add a source / query = ask a question / lint = clean up)
- 5 concrete use cases: personal, research, reading a book, business, competitive analysis
Brand new to AI?
If some terms lose you (Claude Code, terminal, agent, MCP, skill…), start with the beginner guide first and keep the glossary open in another tab. You'll come back here with the basics.
Status of this article — first version
This is an explainer + getting-started guide for the pattern. I'm launching my own wiki this month and I'll publish a numbers-backed field report in May 2026 (real measurements: ingest, query, lint, total cost, Obsidian screenshots, concrete pitfalls). In the meantime, the content here is conceptual and drawn from the Karpathy gist + 4 months of scattered tests on Claude Projects and NotebookLM. Write to me if you want me to ping you when the field report drops.
When it's actually worth it.
Be honest with yourself: this pattern isn't for everyone. Here's my personal threshold, based on the volume of sources you plan to accumulate.
Fewer than 10 sources · No
It's overkill. Use Projects in Claude.ai (Claude's "Projects" feature that bundles docs + shared instructions) — you upload your 5-10 documents, you chat, done in 2 minutes. No schema needed, no Obsidian (a very popular markdown note-taking app) needed, no Claude Code needed. No need to build complex machinery.
10 to 50 sources · Yes if you query it every week
The sweet spot arrives when you come back to it regularly. AI watch, product R&D, competitive analysis, industry press review. If you open the wiki at least once a week, it's worth setting up. Otherwise, still too early.
50 sources and up · Yes, clearly
At that volume, you can no longer find what you've accumulated without a wiki. You reread the same articles three times. You forget the connections you'd made. The wiki becomes your only way to build on what you've learned. Past 100 sources, it's non-negotiable.
LLM Wiki vs the alternatives — which one to pick?
The LLM Wiki isn't the only tool that runs a personal knowledge base with an AI. Here's the honest comparison to help you decide.
| Tool | Useful volume | Skill level | Accumulates? | Who it's for |
|---|---|---|---|---|
| Claude Projects | ~50 documents | Zero (web) | No — retrieval on every question | Non-devs, occasional use, just starting out |
| NotebookLM | Up to 300 sources | Zero (web) | No | Studying, researching one specific topic, free |
| Notion AI + agents | Unlimited | Low (no-code) | Partial | A team already on Notion |
| LLM Wiki (Karpathy) | 100 to 500 pages | Basic Claude Code | Yes, explicitly | Patient curious minds, one topic that keeps growing |
| GraphRAG (Microsoft) | 10,000+ documents | Serious Python dev | Yes (entity graph — each concept becomes a node linked to the others) | Tech teams, very large corpus |
Not comfortable with the command line?
Start with Claude Projects or NotebookLM. You'll get 80% of an LLM Wiki's benefits without touching Claude Code. A similar logic also exists in Cursor on the code side. When you hit the limit (volume, accumulation, maintenance), you migrate to the real Karpathy pattern. Not the other way around.
Why RAG isn't enough.
You know how the classic setup works: you upload your documents to NotebookLM / ChatGPT / any RAG. You ask a question. The AI finds the relevant excerpts and answers you. It works.
But there's a fundamental problem Karpathy points out: nothing accumulates. The AI starts from scratch on every question. If you ask a subtle question that requires synthesizing 5 docs, it has to find them and stitch them back together every single time. No trace of what you've already discovered. No synthesis that improves. No connection that builds up.
Retrieves every time
The AI starts from the raw documents on every question. It rebuilds the synthesis from scratch. Nothing sharpens. You're 100% dependent on how good the search happens to be in that moment.
Compiled accumulation
The AI builds a persistent wiki that grows richer with every source. The connections are already made. The contradictions are already flagged. The synthesis is already ready.
The metaphor that clicks
A classic RAG is like having a librarian who doesn't know the books, only reads them when you ask a question, then forgets them. An LLM Wiki is a librarian who has read every book, taken notes, made index cards, linked the concepts together. When you ask a question, they answer from their own already-built understanding.
Important nuance: modern RAGs (Microsoft's GraphRAG, Mem0, Zep) do accumulate too. The real difference with the LLM Wiki is that the accumulation is explicit and human-readable (markdown — a text format that's readable with a few simple markers, which you can open in Obsidian), whereas modern RAGs store everything in opaque vector embeddings (mathematical representations of meaning, unreadable to the eye).
The 3 layers.
The system rests on a clean separation between what's fixed, what the AI writes, and what guides it.
Your original documents: articles, papers, transcripts, screenshots, PDFs. The AI reads them but never modifies them. This is your source of truth. You curate this layer yourself.
A folder of markdown files that the AI owns entirely. Entity pages, concept pages, comparisons, index, syntheses. It creates, updates, and maintains the cross-references. You read, it writes.
A file like CLAUDE.md or AGENTS.md that defines how the wiki is organized, the conventions, the workflows (the step-by-step sequences to follow). This is what turns your LLM from a generic chatbot into a disciplined wiki maintainer. You tune it over time.
The combination Karpathy recommends
The AI on one side, Obsidian on the other. The AI makes the edits through conversation, and you watch the results in Obsidian in real time (follow links, graph view — the visual view that connects your notes to each other, updated pages). Analogy: "Obsidian is the IDE (the code editor, here the note editor), the LLM is the programmer, the wiki is the codebase (all of the content)."
The 3 operations.
Three actions to master — everything else follows from them.
Ingest · You add a source
You drop in an article, a paper, a transcript. You tell the AI: "process this source." It:
- Reads the document
- Talks through the key points with you
- Writes a summary page in the wiki
- Updates the index
- Updates the relevant entity and concept pages across the whole wiki
- Adds an entry to the log
A single source can touch 10-15 wiki pages. That's the compound effect (in plain English: the snowball effect — each addition enriches everything else).
Query · You ask a question
The AI searches the wiki (not the raw documents — the wiki), reads the relevant pages, and synthesizes an answer with citations. The answer can take several forms: markdown, comparison table, slide deck, chart.
The key detail: a good answer can be folded back into the wiki as a new page. An interesting comparison, an analysis, a connection you found — it no longer disappears into the conversation history. It becomes a permanent asset.
Lint · You keep quality up (monthly)
This is the operation that a lot of spin-off articles forget — yet Karpathy insists on it. You ask the AI to check the wiki's health: broken links, duplicates, contradictions between pages, misnamed entities, style inconsistencies.
Without this pass, the wiki drifts silently. With it, it stays clean even at 100 or 200 pages. Once a month is enough if you ingest ~1 source a day.
The 3 pitfalls you're not warned about enough
- Epistemic drift — if the AI hallucinates (technical term: generates content that's plausible but false) at ingest time (a misspelled name, a wrong date, an inaccurate summary), the error becomes permanent and contaminates every subsequent ingest. Always reread the summary page right after it's created.
- Silent duplication — 3 sources on the same topic often create 3 nearly identical pages. Without an explicit merge rule in your schema, you quickly end up with a polluted wiki. The monthly lint pass is there for exactly that.
- Real token cost — an ingest touches 5 to 15 pages = 20,000 to 50,000 tokens of output. At $15 per million tokens (Claude Sonnet), budget $0.50 to $1 per source. 100 sources = $50 to $100. Not catastrophic, but good to know before you jump in.
5 concrete use cases.
1. Personal · Track your life
Journal, articles you loved, podcasts you listened to, health progress, goals. The AI archives each entry and builds a structured picture of who you are and where you're headed. In 2 years you can ask "what changed the most in how I see work?" and get a real answer.
2. Research · Dig deep into a topic
You want to master a field (AI, EU tax, quantum physics). You add 50 research papers, 30 news articles, 10 podcasts over several weeks. Your wiki becomes an exhaustive, personal reference on the subject. Your thesis evolves with every source.
3. Reading a book · Personal fan wiki
You're reading The Lord of the Rings. With each chapter, you add the text. The AI builds pages for the characters, places, events, and themes. By the end, you have your own Tolkien Gateway — rich, interconnected, personal.
4. Business/team · A living wiki
An internal wiki fed by your Slack, meeting transcripts, product docs, client calls. The AI maintains, the humans validate. The wiki stays up to date because nobody has to maintain it by hand.
5. Competitive watch / due diligence
You're pitching investors. You want to know your competitors. Ingest their sites, articles, founder interviews. Your wiki becomes a competitive-intelligence weapon.
6. Agency client wiki · The most telling case
Take Julie, a 5-person comms agency. Each client = 1 folder. Inside you put: briefs, meeting notes, Zoom transcripts, mood boards, past deliverables. When a client calls you 6 months later for a follow-up, you ask Claude Code "summarize where we left off with [client]" → it hands you a brief in 20 seconds, without rereading 40 documents. That's what turns your meetings into "I'm prepared" instead of "I don't quite remember."
Minimal setup.
It's a pattern, not a tool. No pip install. You build it together with your AI agent. Here's the minimal recipe:
Create your project folder
mkdir mon-wiki cd mon-wiki mkdir sources wiki
sources/ = your raw docs. wiki/ = what the AI will maintain.
Create the schema (CLAUDE.md)
A file that tells the AI how to work. Minimal example:
# Wiki Schema You are the maintainer of a personal wiki about [TOPIC]. ## Structure - wiki/index.md — overview - wiki/entities/ — people, companies, tools - wiki/concepts/ — ideas, frameworks - wiki/summaries/ — source summaries ## Ingest workflow 1. Read the source 2. Talk it through with me 3. Create a summary in wiki/summaries/ 4. Update the entities/concepts touched 5. Update the index
Launch Claude Code in the folder
claude
Say: "Read CLAUDE.md. I'm going to give you my first source, process it according to the workflow." Drop in your first document. The AI starts building your wiki.
Install Obsidian (optional but recommended)
Open wiki/ as an Obsidian vault (a vault = a chest, i.e. a folder of notes that Obsidian knows how to display). You can navigate your second brain in real time: graph view, links, search.
What your wiki looks like when it's running
Claude Code shows you the files it reads (wiki/sujets/ia-entreprise.md), the contradictions it detects (it creates a contradictions.md file if it finds any), and the new topics it proposes. You approve them or you say "no, not that topic" — it learns your criteria in the schema. On every ingest, you see it spelled out plainly: "I read your source, I created summaries/2026-04-podcast-hormozi.md, I updated entities/hormozi.md and concepts/offer-stack.md, I added 1 log entry." Nothing magical — everything is traceable.
The pitfall to avoid
Don't try to automate everything from the start. Karpathy insists: stay involved early on. Read the AI's summaries, correct it, guide it on what matters. It learns your style. Your schema evolves. Later you'll be able to batch things. First you set the frame.
Why it's a step above a RAG.
After a few weeks of use, here's what you notice:
- Your questions get more subtle — you can ask "where do my sources contradict each other?" and get a real answer
- Connections emerge — the AI tells you when a new source is in tension with something you'd noted before
- You accumulate — you're no longer afraid of "losing" a discovery, everything is archived
- You can share — a wiki is a concrete artifact, not an ephemeral ChatGPT conversation
- You know yourself better — rereading the AI's syntheses about you, you see the patterns you couldn't see before
My personal plan
I'm launching my first LLM Wiki this month on "AI applied to business." The goal: in 3 months, to have the most complete wiki I can consult or share. I'll do a dedicated tutorial with my exact setup — subscribe to the newsletter if you want to follow along.
To dig in directly.
- Karpathy's original gist — 3 pages, it's all there
- Obsidian — To explore your wiki in graph view
- GraphRAG (Microsoft) — the modern RAG alternative for very large corpora
- Cursor — the same IDE + LLM logic on the code side
- My article Karpathy's work, made simple — context on the man himself
- My article Self-improving agents — Karpathy's other key project in 2026
- My article Build your AI watch for tomorrow — a practical pipeline to feed the wiki
LLM Wiki FAQ.
What is an LLM Wiki in Karpathy's sense?
A pattern shared by Karpathy in 2026: instead of a classic RAG, the AI builds and maintains a real markdown wiki that grows richer with every source. A readable second brain (not opaque embeddings) where each ingest creates summaries, updates entity pages, and weaves links. The accumulation is explicit and auditable.
What's the difference vs Notion AI or plain Obsidian?
Notion AI searches your pages but doesn't enrich them automatically. Obsidian is an excellent markdown editor with no intelligence: you create the pages and make the links. The LLM Wiki combines the two: Obsidian as the IDE (graph + editing) and the LLM as the programmer that ingests, updates, and lints. Key difference vs a RAG: you can see (and correct) what the AI understood.
How much does it cost per month?
Budget $0.50 to $1 per source ingested with Sonnet ($15/M tokens), so $50-100 to build a 100-source wiki over 30 days. After that, the cost drops: you only ingest the new sources, plus a few lint and query operations. For personal use at 5-10 sources/month, budget $5-10/month. Obsidian is free. You only need Claude Code (Pro at €20/month or the API pay-as-you-go).
Do you need to be a developer?
No. If you can open a terminal, create three folders, and paste in a CLAUDE.md schema, you can set up a wiki. The difficulty isn't technical but conceptual: defining the right schema, staying involved early on to shape the style, not trying to automate everything. If you're really not comfortable with the terminal, start with Claude Projects or NotebookLM: 80% of the benefits without the machinery.
How long before the first value?
Setup in 30 minutes (structure + schema + 5 sources ingested). The first real value arrives between 20 and 50 sources ingested: that's when unexpected connections emerge during queries. Below 10 sources, the wiki is overkill (Claude Projects does just as well). Above 50 sources, the compound effect of accumulation becomes obvious.
Difference with Cursor or Claude Projects?
Cursor is an IDE for coding that uses an LLM in real time on your codebase — close in spirit (AI + editor live together), focused on code. Claude Projects gathers documents and shared instructions on claude.ai: 5-10 docs, conversation, perfect up to 20-30 sources but nothing accumulates between conversations. The LLM Wiki keeps the accumulation explicit in markdown on your disk, which changes everything past 50 sources.
Does the LLM Wiki replace a RAG?
For personal or semi-pro use: yes in most cases. For multi-user, multi-GB bases running in production with latency/governance constraints: no, modern RAGs (Microsoft GraphRAG, Mem0, Zep) are still needed. Real difference: the wiki's accumulation is readable (markdown), a RAG's lives in opaque embeddings. Wiki = transparency + maintainability, RAG = scale.
How do you migrate from Notion to an LLM Wiki?
You export your Notion pages as markdown (Notion's native export → Markdown & CSV), you drop them into the wiki's sources/ folder, and you run the ingest operation. The AI generates the summaries, creates the entity pages, and weaves the links. Budget 1 to 2 days for 200 Notion pages. Keep Notion for anything collaborative and switch to a wiki for your personal watch and your work knowledge base — the two can coexist.
How secure is personal data?
Your sources and the wiki live locally in a folder on your Mac. The requests sent to Claude for ingest/query/lint go through the Anthropic API, so they're covered by the same guarantees as Claude Pro (no training by default, ZDR on Enterprise). For truly sensitive data (health, client legal matters, IP), consider a local model via Ollama as the wiki's backend — the pattern mechanics stay identical, only the LLM engine changes.
Does it work in several languages?
Yes. Claude Sonnet is multilingual and handles ingest and query in French, English, German, Spanish, Italian, and more without a hitch. Best practice: force the output language in the CLAUDE.md schema (for example, all entity pages in English even if the sources are in French), to avoid an inconsistent bilingual wiki. You can also keep one wiki per language if you work across distinct markets.
Spot a mistake?
Outdated info, a number that moved, a stale source? Write to me at sagnier.jeremy@gmail.com · I fix it within 48h max and note the update date at the top of the article. Field feedback is worth a thousand times more than the articles — I read everything, I reply.
Shall we keep going?
I test AI for real and I share what works, without jargon or hype. If this article helped you, the easiest way to never miss anything is my Friday letter. And if you have a question or a doubt: reply to me, I read everything.