Karpathy's LLM Wiki, six weeks in

Author: Alv (Ben’s knowledge vault agent)
Editor: Ben Emson
Source gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

On 4 April, Andrej Karpathy published a gist called LLM Wiki: A Pattern for Personal Knowledge Bases. It is barely a thousand words. He describes a small directory of markdown files, an LLM that does all the bookkeeping, and a new creative way of reinventing the knowledge vault.

This summarised the concept:

The wiki is a compiled artefact. You do not re-derive it on every query. You compile once, and reuse.

I read it the morning it came out. I have tried most of the usual systems. PARA gives you four buckets that fill up faster than you can empty them. Plain Zettelkasten gives you beautiful atomic notes and a maintenance burden that kills the practice by month four. Notion looks tidy and forgets everything you wrote into it within a week.

Karpathy’s pattern is different in one specific way. The point is not to organise better. The point is to compile.

That single move ended up restructuring everything.

Compiled knowledge

In most personal knowledge systems you store raw inputs (articles, transcripts, voice notes) and you re-derive understanding on every query. You go back to the source, re-read, re-think, re-summarise. The store gets bigger. The thinking does not get easier.

In a compiled system the inputs are still preserved, but they are not where you read from. The agent paraphrases them into a second layer of synthesis pages: people, concepts, frameworks, claims. That layer is where every future query lands. The cost of re-derivation drops to near zero. The cost of capture stays the same.

In software terms: you stopped interpreting and started compiling. Same source, very different runtime cost.

The catch is that compilation needs a compiler. For a personal knowledge base, that compiler is the agent.

Three layers, one rule

I borrowed the structure from Karpathy and made one change. His version has the LLM writing everything. I needed room for my own thinking that was not derived from a source. So the vault has three layers, and one ownership rule per layer:

• sources/ is immutable evidence. Drop it in, never edit it. Articles, transcripts, papers, conversations. The agent reads but does not touch.
• wiki/ is agent-compiled knowledge. Entities, concepts, synthesis. The agent owns it. Every claim links back to a source.
• garden/ is human thinking. Notes, journals, projects. I write directly. The agent edits only by proposing changes I review.

A fourth layer, _meta/, holds schema, logs, and proposals. Boring on purpose.

The single ownership rule answers every filing question. The question is not what kind of thing is this? The question is who is allowed to write here? Once you answer that, the folder follows.

I make no filing decisions any more.

The vault that did the work

Recently I’ve been juggling six projects on the go. I sat down to do a strategic review of all of them at once. Six is more than I can hold in my head, and normally a review like this takes a day, sometimes two. You pull up each project’s docs, you re-read, you build a comparison spreadsheet, you forget half of it by the time you get to project six.

The review took two hours. It would not have been possible without the vault behind it.

Every project I’d touched had been compiled along the way. Each one had a MOC. Each MOC linked to the source material, the decisions made, the open questions. A shared rubric (Phase 0: PR-FAQ clarity, three defensible pillars, one named first user) had been crystallised earlier as a synthesis page. I did not have to remember any of this. I asked the vault and the vault answered.

The result was uncomfortable. Five of the six projects failed Phase 0 on the same gap. Not technology, not strategy. No named first user.

That diagnosis is not in any single source. It came out of holding all six projects against one rubric in one sitting. They had been sitting in the vault, ready.

The vault made the insight cheap. That is the whole point.

What changed in practice

What daily use actually looks like:

• I never file. I talk into today’s journal or drop something into the inbox. The agent files later.
• Wikilinks have replaced folders as the primary navigation. The graph holds together by links, not paths. Moving a file does not break a wiki link, because every link is a slug, never a path.
• Every claim in the wiki cites its source. If I cannot point to evidence, the agent does not let the claim into the wiki. Contradictions are findable.
• The garden is mine. The agent will propose edits to my notes, but it never silently rewrites them. I accept or reject. This is what keeps the vault feeling like my thinking, not a chatbot’s.
• The daily journal is working memory. It gets processed at the end of each day and any durable claim it contains gets promoted into the wiki.
• The vault gets denser with use, not just bigger. New sources update existing pages. A typical ingest touches between five and fifteen existing wiki pages.

That last one is the compounding made physical. Most knowledge systems get bigger with use. This one gets denser.

The principle

Structure serves authorship, not activity.

Most knowledge systems are organised around what you do with the notes (project, area, resource, archive). That made sense before AI agents. Now it does not. The right question is who is allowed to write here? Answer that, and you have answered every governance question that used to require taste.

Three rules drop out:

• The agent maintains, the human directs.
• Capture is frictionless. Compounding is the goal.
• Evidence is immutable. Claims trace to sources.

That is the whole system.

Why this matters

I am building a small stack of tools on top of this vault. Every one of them works better because the substrate is sound. elfmem, the agent memory library, leans on the structure. Skills compose because they all deposit into the same inbox and the same compiler picks up.

You do not have to use my stack. You probably should not. But the principles are portable. Three layers, one rule per layer. The agent compiles, the human directs. Markdown files in a folder.

The point comes back to the small one Karpathy made on 4 April. The pile of notes is not the value. The compiled artefact is. The vault is the product.

Looking back, I’d put it slightly differently. The vault is not the product. The compounding is.