Anthony thinks out loud. He works through a problem by talking, and sometimes what he says sounds like a decision. Later it comes back as something the team did, with "well, you said" attached. Chris has watched an offhand remark turn into a company rule four or five times in growing companies.
That's how institutional knowledge goes wrong. Nobody lies. A guess gets written down or repeated, and everyone forgets it was a guess.
In 2020, Oxide shared the rules for how its team writes down ideas. Each document, which the company calls a request for discussion, carries a status: ideation, discussion, published, committed, or abandoned. The label keeps a guess looking like a guess.
Companies have always run on tribal knowledge, the stuff that never made it into an SOP, and they run surprisingly well on it. Anthony's warning is that once AI agents are involved, everything breaks without a solid memory system. So founders write more down. Chris used to capture lessons in a sprint retro. Now he has hundreds of markdown files written for AI context, not for him. That's a local memory. A company needs a shared one, and more files don't make one on their own.
Stephanie defines institutional knowledge as data with meaning. Chris adds the test: can anyone act on it? Startups are supposed to be learning organizations, so the question isn't whether the founder learned something. It's whether the company did.
Then time gets involved. Stephanie had Claude go through ten years of her call recordings, documents, courses, and social posts to find the patterns and build her brain. The catch is that her beliefs changed over those ten years, and a brain built from what she said a decade ago conflicts with what she thinks now. Her new business ran on a different model three months ago, and Claude kept building documents from the old one until she retaught it. AI doesn't forget. Ask it for your address and it holds your home, your office, and the place you lived three years ago.
Anthony's fix: write down why you made a decision and when, because things change. A reason and a date let someone tell a live decision from a retired one. For the spoken version, he now flags his musings with "I'm just thinking out loud. What if..." One sentence keeps a half-formed idea out of the company's memory.
What's worth capturing? Stephanie starts with what you learn about your business and your clients, plus how the team actually works, because those patterns show what the company values, which may not match what it says it values. She's warier of documenting processes early. Codifying a process is easy, and it's a trap: do it once and you stop innovating. The core is smaller: who are we, what are we doing, what do we believe, and what do we need to know that gives us a competitive advantage.
Then comes the part no tool covers. AI can capture and filter what matters, but Chris argues it never replaces a leader standing in front of the team and saying here's what we learned and why, because people still learn best through stories. Stephanie agrees: a machine can't translate, interpret, or say out loud why something matters. Anthony adds that AI will tell you you're right when you're not, and if you could do it with AI, so can somebody else. Documented knowledge with a human in the loop is, in his view, the differentiator.
Start small. Anthony points to Andrej Karpathy's wiki approach, an agentic wiki that builds a company brain in a tool like Obsidian or Notion, and expects a growing company to outgrow it. Just don't let the system become the job. He knows a startup that spent three weeks choosing a CRM before it had a customer.
Write down the why and the when. Mark your guesses as guesses. Then go tell the team the story yourself.
Watch the Full Episode on Building Institutional Knowledge below:
Follow us to watch live on YouTube and LinkedIn or listen to episodes on Apple Podcasts and Spotify.