Essay
Every organisation needs a second brain
Building a personal second brain over 728 desktop files surfaced a bigger pattern: most corporate knowledge is locked in formats AI can't reason over. The next platform shift is AI-native knowledge that compounds.
Last night I spent a few hours building my second brain. A file system that can think and be talked to. 728 files sitting on my desktop — PDFs, PowerPoints, Word docs, spreadsheets — turned into a local knowledge base.
The immediate benefit: information retrieval has never been so relevant and efficient. A semantic layer built on top of those 728 files. 3,359 wikilinks (how files relate to each other by meaning, not filename). Vector embeddings. A hybrid search index. All running locally.
An example — you ask: “What can you find for requirements of people leading engineering teams?” Keyword search, file content matching, and semantic query expansion all fire at the same time. The result: rich content in relevant context, from role descriptions to career frameworks to case assignments. Not just one file. A web of connected knowledge.
My setup was inspired by Karpathy’s LLM Wiki concept — stop treating your documents as dead files you search over every time. Instead, have LLMs build a persistent, interlinked knowledge base that compounds.
The stack
- Obsidian — local, free, no vendor lock-in
- File converters — PPTX / PDF / DOCX / XLSX to Markdown
- QMD — three tiny models running locally for vector embeddings, LLM reranking, and query expansion
- Claude Code — automatically built 3,359 semantic links across all files, like a diligent bookkeeper that never gets tired
Browsing the Obsidian graph view and stumbling on a link between a strategy deck and a research doc from last year — that kind of serendipity doesn’t happen when you’re just prompting an agent.
Why this matters at the enterprise scale
Every person should have a second brain like this. And every organisation needs one too.
Most corporate knowledge is locked in formats AI can’t reason over. PowerPoints that store layout, not meaning. PDFs that are basically screenshots of text. Word documents whose structure encodes nothing about how concepts relate.
The future isn’t better search over bad formats. It’s AI-native formats that compound knowledge over time.
What would it take for organisations to build theirs?
Originally shared on LinkedIn.