Everything is AI. It has been for years. Products, processes, solutions. Without an AI label, apparently nothing is worth anything anymore. And yes, I use AI every day myself—more precisely GenAI and what LLMs and agents can do. Because AI itself has been around for decades. Nobody just noticed. 😉
Even though I generate code with AI or discuss architectures with it, one topic stands out for me: search in large amounts of data. Whether it is a wiki or a collection of URLs, finding the right thing again is always a bit of a gamble. The Zettelkasten tried its own path here, one that is also quite tedious to keep up. With LLMs we now have something at hand that “understands” our content.
A Problem Older Than LLMs
The question is not new. In 1945 Vannevar Bush described the Memex, a store in which documents are connected by associative trails; what he could not solve was who maintains those trails. Luhmann’s Zettelkasten lasted for decades because one man kept a keyword register by hand. Tags and folksonomies rot with the first day you stop handing them out. The Semantic Web did not fail on technology—it failed because someone has to model every class and every relation. Full-text search finds strings, not meaning. And the last attempt before LLMs was semantic search: LSI in 1990, LDA in 2003, Word2Vec in 2013, vector databases today. It finds what is similar, not what is right, and it returns a ranked list, not an answer. Every one of these solutions had to know the question before it was asked.
The Reflex Called RAG
The reflex that arrives immediately is called RAG (Retrieval-Augmented Generation): split sources into chunks, embed
them, fetch the most similar passages, and let the model answer from them. But there is a simpler way, as Karpathy
showed with his LLM Wiki method. He gets by with a
plain index.md and needs no embedding database. What matters is a good essence of the knowledge and a few sensible
agent tools—not the infrastructure behind them.
From Sources to Claims
Thought Keeper had been on my list for a long time. My endless link collections and Markdown files never had a properly searchable home, and with an integrated agent I finally built the thing that makes them pay off.
Thought Keeper takes PDFs (OCR included), Markdown files, or URLs. A small wiki is built in, so I can edit pages right inside the app. From all of that, an agent writes so-called claims—the atoms of the content—clustered by similarity. The claims form a fine-grained Zettelkasten I can search and follow cross-references through. It is the same idea as Karpathy’s, one level deeper: instead of pages, Thought Keeper indexes claims and keeps the source node every sentence came from.
On top of its own data I can also ask questions, which a model and a few tool calls turn into a cleanly written answer. The language of the sources stops mattering, since most models interpret different languages well enough; the answer comes back in the language I asked in. In the end, when I read it, I no longer care which source every sentence came from. I get something that appears in none of my sources as such—an answer that only exists because the sources were combined. I can still look it up whenever I do not trust it. Beyond the hype about code generation and automation, that is one of the most fascinating things LLMs can do for me. And it is not a property of this project: before LLMs, no amount of indexing, tagging or vector search had ever produced an answer that exists in none of the sources.
The Upkeep Never Went Away
The upkeep has not disappeared. It only changed hands—and with it, the kind of failure. A knowledge system that orders itself inherits exactly the work no human ever finished: it has to be true. In Thought Keeper that means provenance, idempotent re-ingest, and an index that can be rebuilt from the files at any time. The question is no longer whether I maintain the order, but whether the machine keeps its own promises.
Further Exploration
- Agentic Coding, Not Vibe Coding: Delegating the Typing, Keeping the Architecture—delegating the typing while keeping the architecture
- A Year of Recalibration: Rethinking the Tools I Use Every Day—where AI sits in my current stack
- Digital Sovereignty: Beyond the Buzzwords—why the vault stays on my own machine
- AI in Software Development: Tool, Not Replacement—a tool, not a replacement
- Beyond the Tool Trap: Rediscovering Real Innovation—tools are not the point