Self-Maintaining Knowledge Vault

Every knowledge base decays the same way. Somebody sets it up carefully, it is accurate for a month, and then reality moves and the document does not. The usual fix is discipline, which is another way of saying it will decay.

So the maintenance is scheduled rather than intended. Nine jobs run across the week, each with one source and one destination. Meeting notes are pulled in and filed. Engineering channels are swept for anything that breaks the local environment, and the runbook is updated when something does. The vault is diffed against a stored index, so a digest can say what was actually added, updated and removed rather than what someone remembers. New research is filed into the right folder.

Filing is only half of it. Every summary document that depended on the old state is brought back into agreement. That is what makes it a system rather than a pile of scripts: material does not wait in an inbox to be sorted, it arrives, it is filed, and whatever summarised the previous state is corrected. It never invents a decision or a number, and anything it inferred is marked as inferred.

It is maintained, not just running. One job is retired with a written note not to turn it back on, because what it did was folded into another. One is paused and run by hand, because it deletes things. One exists only to remind someone to run that one, and reschedules itself. Every job runs on a single machine while the app is open, so a missed run fires on next launch rather than never.

It started as the Course Companion team's vault, and it became the model for the whole company.