LLM Wiki
The LLM wiki: an encyclopedia your AI keeps for you.
Andrej Karpathy made a simple idea popular: instead of letting an AI search from scratch every time you ask, let it write a wiki from your sources and keep it up to date. Here’s why that works, a template you can copy today, and how the script setup compares to a hosted one.
Reading time: 12 minutesUpdated:
01
The idea: the AI writes the wiki, you read it
Picture two ways of handing your knowledge to an AI. In the first, you throw every document onto one big pile. Each time you ask something, the AI rummages through it, pulls out a few paragraphs that look relevant and builds an answer. Tomorrow, with the next question, it starts over. That’s RAG, the usual technique behind “chat with your documents”.
In the second, the AI reads each source once, carefully, and works it into a wiki: one page per source, one page per concept, links in between. When a new article arrives, it updates the pages it touches and notes where it disagrees with older sources. That’s the LLM wiki as Andrej Karpathy described it.
The setup has three layers. Raw sources stay untouched. The wiki next to them is plain Markdown that the agent writes and you read, in Obsidian for example. And a schema, one file of rules (in Claude Code that would be a CLAUDE.md), tells the agent how the wiki is organised. Day to day, you do three things with it: ingest a new source, query the wiki, and now and then lint it for gaps and contradictions.
02
RAG or LLM wiki?
Both give an AI access to what you know. The difference is when the thinking happens.
| – | RAG: search on every question | LLM wiki: work it in once, keep it current |
|---|---|---|
| When the work happens | On every question, again and again | Once, when new material arrives |
| What the AI reads | Snippets from raw documents | Maintained pages with a summary, sources and links |
| Connections across sources | Rebuilt from scratch every time | Already on the page |
| Contradictions | Rarely noticed | Written down while working a source in |
| What you can read yourself | An index made for machines | A wiki you can browse |
| Best for | Huge piles of documents nobody curates | Your own knowledge and topics you go deep on |
- When the work happens
- RAG: search on every questionOn every question, again and again
- LLM wiki: work it in once, keep it currentOnce, when new material arrives
- What the AI reads
- RAG: search on every questionSnippets from raw documents
- LLM wiki: work it in once, keep it currentMaintained pages with a summary, sources and links
- Connections across sources
- RAG: search on every questionRebuilt from scratch every time
- LLM wiki: work it in once, keep it currentAlready on the page
- Contradictions
- RAG: search on every questionRarely noticed
- LLM wiki: work it in once, keep it currentWritten down while working a source in
- What you can read yourself
- RAG: search on every questionAn index made for machines
- LLM wiki: work it in once, keep it currentA wiki you can browse
- Best for
- RAG: search on every questionHuge piles of documents nobody curates
- LLM wiki: work it in once, keep it currentYour own knowledge and topics you go deep on
A well-kept wiki doesn’t need fancy retrieval. WhyVault uses full-text search – because the pages are clearly named and linked, that’s enough for the agent to find what it needs.
03
What people build with an LLM wiki
Anything you keep reading about, and want to understand rather than just store. Six examples:
A topic you go deep on
Sleep, nutrition, AI agents, a period of history. Twenty articles turn into ten concept pages, each with sources and open questions, instead of twenty bookmarks.
Books and courses
One source page per book or module, concepts across all of them. A year later you know how the books connect, not just that you read them.
Market and competitors
Paste in pricing pages, reviews and launch posts. The agent keeps one page per competitor and a map of where the market is heading.
Your team’s know-how
Customer calls, post-mortems and meeting notes become pages your next hire, or your next agent, can read in ten minutes.
Podcasts and talks
Drop in a transcript and ask for the source page. The ideas land on the concept pages they belong to, linked to the minute they came up.
Research for a decision
Which CRM, which city to move to, which laptop to buy. Collect sources, let the agent compare, and keep the conclusion with its reasons.
04
Why it works so well
Keeping a wiki was always a good idea. What’s new is that someone else does the tedious half.
Knowledge compounds
Every new source makes existing pages better instead of adding one more file to the pile. After a few weeks, one page holds more than any single document.
The agent does the upkeep
Summarising, linking, updating cross-references: exactly the maintenance where every second brain eventually stalls. An AI doesn’t get bored of it.
Better answers
Reading a curated page beats piecing together snippets. That’s true for people, and it’s true for agents.
Your Markdown
The wiki is plain
.mdfiles with[[wikilinks]]. No vendor format, no database you can’t get out of.
05
The script setup: Karpathy’s way, on your laptop
The home-built version is quick to describe: a folder on your laptop with raw/ for sources and wiki/ for the pages, Obsidian for reading, Claude Code or another agent in the terminal, and a CLAUDE.md with the rules. Add the Obsidian Web Clipper to save articles as Markdown, Git to see what changed, and maybe a local search tool once the wiki grows. It works well – for one person on one machine.
Then it gets tricky. Claude on your phone and ChatGPT in the browser can’t reach that folder. Every tool wants its own rules file, so you keep CLAUDE.md, AGENTS.md and your Cursor rules in sync by hand. What the agent changed across twelve pages yesterday, you only see if you committed to Git first. And nothing stops it from overwriting or deleting a page.
Then come the little scripts that collect sources, rename them and file them in the right folder. With every script, your AI second brain turns a little more into a project you have to maintain.
06
Script setup or hosted?
Same idea, same Markdown. The difference is who looks after the plumbing, and where your wiki can be reached.
| – | Script setup (Karpathy-style, local) | Hosted (WhyVault) |
|---|---|---|
| What you need | A folder, Obsidian, Git and an agent in the terminal, such as Claude Code | An account and an agent that speaks MCP |
| Setup | An afternoon, plus scripts as you go | About ten minutes |
| Rules file | CLAUDE.md, AGENTS.md, Cursor rules – one per tool | One note in 9_System, read by every agent |
| Reachable from | The computer the folder lives on | Claude on web, desktop and phone, ChatGPT, Cursor, Claude Code |
| Seeing what changed | Git, if you commit before every run | Automatic history per note, with the agent’s name |
| Undo | git revert or git checkout | One click, or everything one agent did since a point in time |
| Deleting | The agent can delete any file it can reach | Off by default; if you allow it, only into the trash |
| Search | Whatever you set up: grep, a local search engine, embeddings | Full-text search built in, no vector search |
| Freedom to customise | Unlimited: your own scripts, plugins and tools | Fixed tools, no scripts, no background jobs |
| Cost | Free, apart from your AI plan | Free up to 20 notes, Pro €9.99 a month |
- What you need
- Script setup (Karpathy-style, local)A folder, Obsidian, Git and an agent in the terminal, such as Claude Code
- Hosted (WhyVault)An account and an agent that speaks MCP
- Setup
- Script setup (Karpathy-style, local)An afternoon, plus scripts as you go
- Hosted (WhyVault)About ten minutes
- Rules file
- Script setup (Karpathy-style, local)
CLAUDE.md,AGENTS.md, Cursor rules – one per tool - Hosted (WhyVault)One note in
9_System, read by every agent - Reachable from
- Script setup (Karpathy-style, local)The computer the folder lives on
- Hosted (WhyVault)Claude on web, desktop and phone, ChatGPT, Cursor, Claude Code
- Seeing what changed
- Script setup (Karpathy-style, local)Git, if you commit before every run
- Hosted (WhyVault)Automatic history per note, with the agent’s name
- Undo
- Script setup (Karpathy-style, local)
git revertorgit checkout - Hosted (WhyVault)One click, or everything one agent did since a point in time
- Deleting
- Script setup (Karpathy-style, local)The agent can delete any file it can reach
- Hosted (WhyVault)Off by default; if you allow it, only into the trash
- Search
- Script setup (Karpathy-style, local)Whatever you set up: grep, a local search engine, embeddings
- Hosted (WhyVault)Full-text search built in, no vector search
- Freedom to customise
- Script setup (Karpathy-style, local)Unlimited: your own scripts, plugins and tools
- Hosted (WhyVault)Fixed tools, no scripts, no background jobs
- Cost
- Script setup (Karpathy-style, local)Free, apart from your AI plan
- Hosted (WhyVault)Free up to 20 notes, Pro €9.99 a month
Love your terminal and work on one machine? Karpathy’s setup is great, and the template below works there too: just rename the folders. Want the same wiki in every app and on your phone, without maintaining scripts? Hosted is simpler. Both are plain Markdown, so you can switch in either direction with a zip file.
07
An LLM wiki in WhyVault, already set up
The folders of the WhyVault template map almost one to one onto Karpathy’s setup. You don’t build anything – you just add one rules note.
0_Inbox/ raw sources: articles, notes, transcripts
1_Knowledge/ the wiki – flat, shared, kept by the agent
Maps/
Learning.md the index: what exists, what’s missing
Notes/
Spaced repetition.md
Retrieval practice.md
Sources/
Talk – How memory works.md
People/
9_System/
Identity/ USER.md, SOUL.md, IDENTITY.md, CONTEXTS.md
Wiki rules.md the schema: your CLAUDE.md for every agent0_Inbox/- Your raw sources land here – Karpathy’s
raw/. The agent reads them but doesn’t change them, and as long as you don’t allow an agent to delete, every original stays put until you tidy up. 1_Knowledge/- The wiki itself. Flat and shared: one idea gets exactly one page, whichever source it came from.
Sources/holds one page per source,Notes/one per concept. 1_Knowledge/Maps/- A map is the entry point to a topic: which pages exist and what’s still missing. It’s the
index.mdthe agent reads first. The log Karpathy keeps inlog.mdis the history here, written automatically. 9_System/Wiki rules.md- Everything in
9_Systemis handed to every connected agent at the start viavault_overview. So a rules note here works like aCLAUDE.md– except Claude, ChatGPT and Cursor all read it. More on the structure
08
The template: copy, paste, start
Five building blocks, each ready to copy. Paste them into WhyVault, or into a local folder if you prefer the script setup.
- 1
The folders
In WhyVault, the template already has them. Locally, create the same folders, or keep
raw/andwiki/and adjust the paths in the rules.0_Inbox/ raw sources – the agent never changes them 1_Knowledge/ Maps/ one map per topic: the index Notes/ one page per concept Sources/ one page per source People/ one page per person who matters 9_System/ Wiki rules.md the schema every agent reads first
- 2
The rules note
Save it as
9_System/Wiki rules.md. It’s the heart of the wiki: short beats complete – sharpen it whenever the agent does something you don’t like.# Wiki rules You maintain my wiki. I collect sources and read; you write. ## Folders - 0_Inbox: raw sources. Read them, never change them. - 1_Knowledge/Sources: one page per source, named “Type – Title”. - 1_Knowledge/Notes: one page per concept. If it exists, extend it. Never create a second page for the same idea. - 1_Knowledge/People: one page per person who matters. - 1_Knowledge/Maps: one map per topic. It is the index. Keep it current. ## Ingest (when I say “work in the inbox”) 1. Read the whole source before you write anything. 2. Create its source page: three-sentence summary, key points, quotes. 3. Update every concept page it touches. Create pages for new concepts. 4. Link every claim to its source with [[Title]]. 5. Contradictions go under “Where they disagree”. Don’t smooth them over. 6. Add new pages to the map. 7. Tell me which pages you created or changed. ## Query - Answer from the wiki first and name the pages you used. - If the wiki has no answer, say so before you use general knowledge. - If an answer is worth keeping, offer to save it as a page. ## Style - Answer first: the opening two sentences say what it’s about. - Short pages, plain language, one idea per page. - Don’t move or delete anything. Tell me what to tidy up.
- 3
A source page
One per article, talk, book or transcript. The agent fills it in; the frontmatter makes it easy to find later.
--- type: source kind: talk author: date: link: processed: 2026-10-12 --- # Talk – How memory works In three sentences: … ## Key points - … ## Worth quoting > … ## Feeds into - [[Spaced repetition]] - [[Retrieval practice]]
- 4
A concept page
One per idea. This is where knowledge compounds: every new source adds a line, a link or a contradiction. A filled-in example follows further down.
--- type: note sources: [] updated: --- # Concept name Two sentences: what it is and why it matters. ## What the sources say - Claim. [[Source]] ## Where they disagree - … ## Related - [[Other concept]] - [[Map]]
- 5
A map
The entry point to a topic. The agent reads it first, and you see at a glance what’s there and what’s missing.
# Learning What this wiki knows about learning, and what’s still missing. ## Concepts - [[Spaced repetition]] – review at growing intervals - [[Retrieval practice]] – recall instead of rereading ## Sources - [[Talk – How memory works]] - [[Podcast – Learning with flashcards]] ## Open questions - Does mixing topics help outside of maths?
09
How to start your LLM wiki in WhyVault
From an empty vault to your first source worked in, without writing any code:
- 1
Create a vault and connect your agent
Sign up, pick the WhyVault template and connect your agent using the address
https://mcp.whyvault.io. Guides for every agent - 2
Paste your wiki rules
Create the note
9_System/Wiki rules.mdand paste the rules from the template above. Or ask your agent: “Create the note 9_System/Wiki rules.md with this text.” - 3
Drop sources into the inbox
Paste articles, meeting notes or transcripts as notes in
0_Inbox. Or tell your agent in the chat: “Save this text as a note in the inbox.” - 4
Ask the agent to work them in
Use the ingest prompt below. The agent reads your rules first, then the sources, and creates new pages or extends existing ones.
- 5
Check the history
Every change shows up in the history with the agent’s name, as a diff against the previous version. If something’s off, undo it with one click. More on history
- 6
Ask questions
Now your agent answers from maintained pages instead of snippets. And when an answer is good, have it saved as a page of its own – that way your questions grow the wiki too.
10
Three prompts for every day: ingest, query, lint
Karpathy’s three operations as prompts you can paste into any connected agent.
- 1
Ingest: work in new sources
Whenever something new is in the inbox. One source at a time gives the best pages; a handful at once is fine for short ones.
Read the wiki rules and work the new notes in 0_Inbox into the wiki. Then list the pages you created or changed, and anything that contradicts what was there.
- 2
Query: ask, and keep good answers
The agent answers from the wiki and says which pages it used. Good answers become pages, so your questions grow the wiki.
Answer from my wiki: what’s the difference between spaced repetition and retrieval practice? Name the pages you used. If the answer is worth keeping, save it as a page in 1_Knowledge/Notes.
- 3
Lint: the health check
Once a month or so. Ask for a list first, then decide what to fix – the checklist below shows what to look for.
Check my wiki and give me a list, don’t change anything yet: pages nothing links to, contradictions that aren’t noted, concepts without a page, claims without a source, and maps that are out of date.
---type: notesources: - "[[Talk – How memory works]]" - "[[Podcast – Learning with flashcards]]"updated: 2026-09-28---# Spaced repetitionIf you review something at intervals, you keep it longer than after one long session. The gaps can grow as it sticks.## What the sources say- Each recall just before you’d forget strengthens the memory. [[Talk – How memory works]]- In daily life, that means: few cards, every day, keep it short. [[Podcast – Learning with flashcards]]## Where they disagree- The podcast relies on an app that schedules each card. The talk says a simple plan works just as well. Open.## Related- [[Retrieval practice]] – works best when you actively recall instead of rereading.- [[Learning]]
11
What a wiki page looks like
The concept page from the template, after the agent has worked in two sources:
Answer first: the opening two sentences say what it’s about. Read only those and you know enough.
Every claim has a source. The link leads to the source page, and from there to the original in the inbox.
Contradictions stay visible. The agent doesn’t smooth them over, it writes them down. That’s exactly where your own thinking starts.
Links make the web. Through
[[Retrieval practice]]and the map[[Learning]], the next agent finds the connection without having to search.
12
What it feels like
After a talk, with your notes already in the inbox:
You: I put my notes from the memory talk in the inbox. Work them into the wiki.
- Tool: vault_overview
Claude: I’ve read your wiki rules. The notes are mostly about spaced repetition and retrieval practice – there are already pages for both.
- Tool: create_note · “Talk – How memory works”
Claude: Source page created, with the three points the talk put the most weight on.
- Tool: update_section · “What the sources say”
Claude: I’ve extended “Spaced repetition”. One thing stood out: the talk disagrees with the podcast about fixed intervals. That’s now under “Where they disagree”.
You: So what’s the difference between spacing and retrieval?
- Tool: read_note · “Retrieval practice”
Claude: Spacing is about when you study: at intervals. Retrieval is about how: from memory instead of from the page. The two pages link to each other. Want me to save that distinction as a page of its own?
13
Once a month: the tidy-up pass
Even a good wiki grows wild. Every now and then, ask your agent to look through it and tell you:
- which pages link to nothing and nothing links to (that’s what
get_backlinksis for), - where two pages contradict each other without it being noted,
- which concepts come up often but don’t have a page yet,
- which claims have no source,
- whether the maps still show what exists,
- which sources in the inbox are already processed. You do the tidying – agents can’t move anything.
Don’t ask the whole library every time. Have someone write the encyclopedia.
?
Frequently asked questions
What is an LLM wiki?
A wiki of Markdown pages that an AI writes from your sources and keeps up to date. You collect material and read; the agent summarises, links and extends existing pages when something new comes in.
What does Andrej Karpathy have to do with it?
Karpathy described the pattern and made it popular: raw sources, an AI-maintained wiki next to them, and a rules file that explains the setup to the agent. WhyVault isn’t an official implementation – it’s a place where exactly that setup works without your own scripts.
Is there an LLM wiki template I can copy?
Yes, the template on this page: folders, a rules note, and page templates for sources, concepts and maps. It works in WhyVault and in a local folder with Claude Code. The full WhyVault vault structure is also free as an Obsidian template.
Do I need Obsidian for this?
No. You read and edit your wiki in the WhyVault app. If you like Obsidian: the export is a zip that Obsidian opens directly as a vault. Obsidian and WhyVault
Can I keep using Claude Code with my wiki?
Yes. Connect Claude Code to WhyVault with one command, and it reads the same rules note as Claude in the browser or ChatGPT. Your wiki then no longer depends on the folder you started it in. Claude Code setup
Does WhyVault maintain the wiki automatically in the background?
No. Your agent does the work when you ask it to – in Claude, ChatGPT, Cursor or Claude Code. If you set up scheduled tasks there, it can do it regularly; WhyVault itself has no built-in AI.
Isn’t this just RAG?
Not quite. RAG searches raw documents on every question; an LLM wiki works knowledge in once and maintains it. WhyVault doesn’t use vector search, it uses full-text search – and with a well-kept wiki of clear pages, that works well.
How much does it cost?
WhyVault is free up to 20 notes – enough to try the idea with a few sources. A growing wiki needs Pro at €9.99 a month or €99 a year, with no limit. Reading and exporting always work, even without a subscription.
Keep reading
All guidesClaude memory
Built-in memory, Projects or a shared vault over MCP: an honest comparison, and when each one is enough.
ReadObsidian & Claude
How to connect your Obsidian vault to Claude, ChatGPT and Cursor via MCP, locally or in the cloud.
ReadMCP memory server
What a memory server for Claude, ChatGPT and Cursor does, local vs hosted, and what to look for before you pick one.
Read
Your wiki, organised from day one.
Create your vault, paste the wiki rules and let Claude work in your first source. Free, no credit card.