Institutional Memory Is a Feature You Have to Fund

Every organisation loses knowledge at a steady rate. People leave, projects end, the reasoning behind a decision fades until only the decision remains. Nobody experiences this as an event. It shows up as a team rebuilding something that already exists, relitigating a settled question, or repeating a mistake documented nowhere because the person who made it last time has moved on.

What is actually lost

Not the code, which is in the repository, and not the documents, which are in the drive. What disappears is the reasoning: why this approach rather than the obvious alternative, what was tried and failed, which constraint made the elegant solution impossible. Without that, successors see an odd decision, assume incompetence, and rebuild toward the alternative that was already rejected for reasons they cannot see.

Cheap practices that work

• Decision records. A page per significant decision: context, options, choice, consequences. Fifteen minutes each, and they are the single highest-return document type in software.
• Post-incident writeups without blame. The value is the causal chain, and blame guarantees the chain gets sanitised.
• Exit documentation as standard. When someone leaves, their last week produces a written map of what they held that nobody else does.
• Dated documents. A page with no date is a page nobody can trust, and untrusted documentation is unread documentation.

The reason it does not happen

Writing things down has a cost now and a benefit later, usually to somebody else. Every incentive in a delivery-focused organisation points away from it. This is why it has to be structural rather than cultural: a template that appears automatically, a checklist item in a process, a review that expects the record to exist. Relying on individual conscientiousness produces documentation from the three people who like writing and nothing from anyone else.

The cost of forgetting is paid by whoever arrives next, which is exactly why it never gets prioritised by whoever is here now.

Memory has a shelf life too

The opposite failure is a wiki with four thousand pages, most describing a system decommissioned in 2022. Volume without maintenance destroys trust as effectively as absence. Archive aggressively, mark pages as historical rather than deleting the reasoning, and keep the live set small enough that someone will actually read it. A hundred current pages beat four thousand of unknown vintage.