How I Use Andrej Karpathy's LLM Wiki Setup in My Obsidian Second Brain

The short answer: I use Andrej Karpathy’s “LLM wiki” pattern to run my Obsidian vault as a second brain that an AI maintains for me. The idea is simple. Keep your knowledge in plain markdown, write a schema file that tells the AI how the vault is organized, and let the AI read new sources, summarize them, link them, and flag contradictions. I do the sourcing and the asking. The AI does the filing. I built my vault after a stroke took the memory I used to rely on, and this pattern is what turned it from storage into something closer to a working mind.

What Karpathy’s LLM Wiki Actually Is

Recently, Andrej Karpathy, a founding member of OpenAI and former senior director of AI at Tesla, published a short idea file describing a pattern he calls an LLM wiki.

Here is the cleanest definition I can give: an LLM wiki is a knowledge base the AI builds and maintains, not just one it searches.

Most AI-and-documents setups work like RAG. You upload files, the model retrieves chunks at query time, and it answers. However, the problem is that nothing accumulates. As a result, the model rediscovers your knowledge from scratch on every question.

Karpathy’s pattern is different. The AI reads each new source once, then writes it into a growing set of linked markdown pages. Cross-references get added, and contradictions get flagged. As a result, the synthesis already reflects everything you have read. As he puts it, Obsidian is the IDE, the AI is the programmer, and the wiki is the codebase.

Ultimately, that single reframe is what I had been missing.

Why This Mattered to Me

I am not a productivity hobbyist. I had a stroke, and it changed how my memory works.

The deficit nobody sees is retrieval. You know that you know something, and it will not surface. At first, I tried to fix that with longer checklists. It did not work, because effort was never the problem. Ultimately, the problem was infrastructure.

So I built an Obsidian vault to hold what my brain could not. At first, it worked as storage. However, I was still the one doing all the filing, the linking, and the remembering where things lived. That is the exact work a changed brain is worst at, the same executive-function load I have written about before. By contrast, Karpathy’s pattern moves that work to the AI. Storage I could already do. Maintenance, on the other hand, was the part I needed handed off.

The Three Layers of My LLM Wiki

Karpathy describes three layers. Here, then, is how they map onto my vault.

Raw sources. The inputs: articles, transcripts, book chapters, journal entries, my own daily logs. These stay immutable. The AI reads them and never edits them. This is the source of truth.
The wiki. The markdown pages the AI writes and owns: summaries, concept pages, people, an index. This is most of my Obsidian vault. I read it. The AI writes it.
The schema. A single CLAUDE.md file at the root that tells the AI how the vault is organized, what the conventions are, and what to do when it ingests a source or answers a question. This is the file that turns a generic chatbot into a disciplined librarian. I co-evolve it over time as I learn what works.

If you take one thing from this post, take the schema file. It is the difference between an AI that guesses and an AI that follows your house rules.

How the Loop Works

You curate a source and drop it in, or ask a question.
Ingest: the AI reads the immutable raw source and writes it into the wiki.
The wiki (my Obsidian vault) holds the AI-written, cross-linked pages.
Query: you ask, the AI answers with citations, and good answers get filed back into the wiki.
Lint: the AI health-checks the wiki for contradictions, orphans, and gaps.
A CLAUDE.md schema file governs all three steps so the AI follows your house rules.

The Three LLM Wiki Operations I Run

Karpathy frames the daily use as three verbs. In practice, mine run like this.

Ingest

I drop a source in and ask the AI to process it. In practice, it reads the source, tells me the key takeaways, writes a summary page, updates the index, and revises any existing pages the new material touches. For example, one article can update ten or fifteen pages in a single pass. Meanwhile, I stay involved and read the summaries as they land.

Query

I ask the vault a question. The AI reads the index first, finds the relevant pages, and answers with citations to my own notes. Importantly, the key habit here is that a good answer gets filed back into the wiki as a new page. A comparison, a connection, an analysis. In other words, it should not vanish into a chat history. Over time, my explorations compound the same way my sources do.

Lint

Once in a while I ask the AI to health-check the vault. For instance, it looks for contradictions between pages, claims a newer source has overturned, orphan notes with no links, and concepts I mention often but never gave their own page. This is the maintenance humans always abandon. As a result, because the AI does not get bored, it actually happens.

What This Looks Like on a Normal Day

This is not theory for me. In fact, my vault runs a version of this every day.

For example, an agent I call Henry, the backbone of the system I described in My Brain Has a Backup, does an autonomous pass across notes I touched recently and fires links between them. A recent entry: it scanned about fifteen notes, connected my page on ego to the published piece about how the stroke crushed it, tied together two separate notes about the gap between knowing and doing, and added reciprocal links to each. No new pages where a home already existed. That is ingest and lint running on their own.

My daily note works as Karpathy’s log.md: an append-only, dated record of what happened. My index and map-of-content notes work as his index.md. The morning brief and the council of advisors I run are query outputs, filed back as pages instead of lost in a thread.

The pattern did not replace my setup. It gave structure to the half I had been doing by hand.

What Stays Human

The AI does the bookkeeping. It does not do the thinking.

My job is to choose what to read, direct the analysis, ask the real questions, and decide what it all means. In short, the AI removes the grunt work so the judgment has room to happen. That division is the whole point. After all, the tedious part of a knowledge base was never the reading. Rather, it was the upkeep. Therefore, hand off the upkeep, and keep the meaning.

LLM Wiki Mistakes to Avoid

Skipping the schema file. Without it the AI invents a new structure every session and the vault turns to noise. Write the rules down once.
Treating it like storage. If you only file notes and never ask the AI to maintain and surface them, you have rebuilt a filing cabinet. The maintenance step is the point.
Dumping a hundred sources unsupervised at the start. Ingest a few at a time and read what it writes. You are teaching it your house style.
Writing the wiki yourself. In this pattern you rarely write the pages. If you are doing the filing by hand, the system is not working yet.

LLM Wiki Quick Summary

An LLM wiki is a knowledge base the AI maintains, not just searches. It compounds instead of resetting.
Three layers: raw sources (immutable), the wiki (the AI writes it), and a schema file (your house rules).
Three operations: ingest new sources, query the vault, and lint it for contradictions and gaps.
I run this inside Obsidian with Claude. An agent links my notes daily, my daily note is the log, my index is the map.
You curate and ask. The AI files and maintains. The thinking stays yours.

FAQ

What is Andrej Karpathy’s LLM wiki? It is a pattern for personal knowledge bases where an AI incrementally builds and maintains a set of linked markdown files. Instead of retrieving from raw documents on every question, the AI reads each source once, writes it into the wiki, and keeps the whole thing cross-referenced and current.

How is it different from RAG or ChatGPT file uploads? RAG retrieves chunks at query time and builds nothing that lasts. The LLM wiki produces a persistent, compounding artifact. The links are already there, the contradictions already flagged, and every new source makes the whole base richer.

Do I need Obsidian to do this? No. The wiki is just a folder of markdown files, so any editor works. I use Obsidian because the graph view, backlinks, and plugins make a growing vault easy to navigate, and because the files stay local and mine.

How I Use Andrej Karpathy’s LLM Wiki Setup in My Obsidian Second Brain