On April 4, Andrej Karpathy posted a gist called "LLM Wiki." It's a pattern for building personal knowledge bases that don't suck. Within 48 hours it had 5,000 stars and a comment thread full of people building implementations.
I read it on Sunday morning. By the second paragraph I was sitting up straighter.
Karpathy is describing the same architecture I've been building at GroundTruth OS for the last six months. He arrived at it from the LLM agent side, working on personal knowledge management. I arrived at it from the business operations side, watching AI agents make the same mistakes over and over because nobody had given them a system of record. We converged on the same answer.
That convergence is the most interesting thing that has happened to GroundTruth OS since I started building it.
The pattern Karpathy described
Karpathy's core argument: most LLM-with-documents experiences look like RAG. You upload a pile of files, the model retrieves relevant chunks at query time, generates an answer, forgets everything. Nothing accumulates. Ask a question that requires synthesizing five sources and the model rebuilds the synthesis from scratch every time. NotebookLM works this way. Most enterprise RAG works this way.
His alternative: an LLM-maintained wiki. A persistent, structured, interlinked collection of markdown files that sits between you and the raw sources. When you add a new source, the LLM doesn't just index it. It reads it, extracts the key information, and integrates it into the wiki. Updates entity pages. Revises topic summaries. Flags contradictions. Strengthens or challenges the evolving synthesis.
The wiki is a compounding artifact. The cross-references are already there. The contradictions have already been flagged. Every new source makes the wiki richer. The human curates and asks questions. The LLM does the bookkeeping that humans abandon every wiki over.
He proposes a three-layer architecture. Raw sources (immutable). The wiki (LLM-maintained). The schema (a config file that tells the LLM how to behave). Operations are ingest, query, and lint.
Read it again with GroundTruth OS in mind. Tell me that's not the same architecture.
The pattern GroundTruth OS describes
GroundTruth OS treats business decisions and context the way Karpathy treats personal knowledge. Three layers: source documents (immutable), a canonical layer the AI maintains, and a constitution that governs how the system behaves. When a new decision is made, it doesn't live in a Slack thread or someone's memory. It gets ingested, structured, cross-referenced against existing decisions, and integrated into the canonical layer.
The pitch I've been using on LinkedIn is "run your business like a codebase." Karpathy's pitch is "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase." Different audiences, identical insight.
Both of us reject the rediscover-everything-every-query model. Both of us treat the canonical layer as a compounding artifact. Both of us recognize that the failure mode of every documentation system in history is human maintenance burden, and that LLMs change the math. Both of us reach for software engineering as the operating metaphor because that's the only model where this kind of discipline already works.
We are building the same thing for two different domains. He's building it for personal knowledge. I'm building it for business operations. Same architecture, same insight, same bet.
Why the convergence matters
When two people working in different domains arrive at the same architecture independently, that's a signal. Not proof of correctness. But a signal that the architecture is responding to a real constraint, not a stylistic preference.
The constraint is this: LLMs are good enough to do the work, but they are terrible at remembering context across sessions. The fix is not bigger context windows or more clever prompting. The fix is a persistent, structured layer that survives between sessions and that the LLM is responsible for maintaining.
Karpathy frames this as the next step beyond RAG. I frame it as the next step beyond the PRD. We're looking at the same thing from different sides of the same problem.
And his framing is going to reach a much bigger audience than mine. Five thousand stars in 48 hours. People are building implementations of his pattern in real time. The market is being educated about the architecture I've been quietly building. That's a tailwind.
Where Karpathy is ahead
I'm not going to pretend the convergence is total. It isn't. Karpathy's thinking is ahead of mine in several specific ways, and reading his gist has already given me a roadmap of things to add.
The biggest one: he treats the lint operation as a first-class citizen. Periodic health checks of the wiki. Look for contradictions. Look for stale claims. Look for orphan pages. Look for missing cross-references. GroundTruth OS doesn't have this yet. It should.
The second one: he separates the index (content-oriented navigation) from the log (chronological history). I've been collapsing those into one structure. Splitting them is cleaner.
The third one is the sharpest. He says good answers to queries can be filed back into the wiki as new pages. Explorations compound. The Q&A you do with the system becomes content in the system. I have not been treating exploration as a source of canonical content. I should be.
There are two or three more. I'll write about them as I build them.
What I'm taking from this
Two things, mainly.
First, the architectural bet is right. When Karpathy independently arrives at the same three-layer pattern from a completely different starting point, the odds that the pattern is correct go up significantly. Not certain. But the prior just shifted.
Second, the work I'm doing is part of a larger conversation that's now happening in public. Karpathy's gist will accelerate the timeline for the entire category. I need to ship faster, write more, and stop worrying that the idea is too early. It's not too early. It's exactly on time.
If you're building anything that depends on LLMs maintaining context across sessions, read his gist. Then come back here and read what I'm building. They're the same idea. Different domain. Same bet.