Dashboards hold what can be quantified. The knowledge that runs companies lives in heads, and it surfaces mid-work. That is where to capture it.

What structure misses. The last note argued for structured context: atoms and molecules, definitions an AI can reference instead of guess at. That covers the knowledge you can write down. It misses the kind that runs companies.

Every team has knowledge that lives only in heads. Why a certain kind of account churns. The discount rule nobody wrote down. The metric everyone watches that is quietly misleading, for a reason one person remembers. Dashboards and semantic layers are lossy by construction. They hold what can be quantified and drop the rest.

Why extraction fails. The usual fix is extraction: interviews, workshops, a knowledge-capture project. It does not work. The sessions are heavy, and their output starts to rot the day they end. Tacit knowledge does not surface on a calendar. It surfaces in the middle of the work.

So capture it there. Ambiently, in small pieces. During onboarding conversations. Through light questions in the flow of work: you keep pulling this segment, what are you watching for? A sentence of human judgment gets stored next to the computed number it explains, in the same chunk, with a source and a date. Not a project with a finish line. An accumulation.

Two rules keep it honest. The AI proposes, a human confirms. Nothing gets redefined silently. And the memory stays open: you can read it, see where each piece came from, edit it, delete it. The opposite of the background memory most agent products keep about you.

The open problems are real. Ask too often and the questions get dismissed like a cookie banner. Tacit knowledge is hard to say out loud on demand. And captured knowledge decays too. This is the short version. The full argument is in the original post: The Things We Just Know.