Tribal knowledge is the undocumented understanding of how a system really works that lives only in the memories of experienced team members — the workarounds, the reasons behind odd-looking code, the "don’t touch that" warnings that were never written down. It is real and valuable, but it is invisible, unsearchable, and bound to specific people.
The cost of tribal knowledge is paid at exactly the wrong moments: when someone leaves and takes a load-bearing piece of understanding with them, when a new hire re-derives something the team has known for years, or when an AI agent starts a fresh session with none of the unwritten context and confidently breaks an unstated rule. The knowledge was available the whole time — just not anywhere the next actor could reach it.
Turning tribal knowledge into shared knowledge means capturing those discoveries where they can be retrieved by the people and agents who need them, anchored to the code they concern. The aim is not to document everything, which never survives contact with a deadline, but to persist the durable, hard-won facts — the gotchas, decisions, and invariants — so they outlive the session and the person who first learned them.