Six patterns we see again and again. Each is a symptom of the same missing thing — governed institutional memory — and each points at what to build instead.
Search finds where something was written. Memory knows what is true now, what it replaced, and who is allowed to see it. A ranked list of documents is not an answer — and it has no idea which document is current.
Memory before retrieval — the current, cited fact, not the top ten links.
A wiki is only as fresh as the last human who remembered to update it. Reality moves faster than curation, so the page that looks authoritative is often quietly wrong — with nothing to signal it.
Memory supersedes automatically and keeps the history; staleness is flagged, not served.
Retrieval-augmented generation embeds text chunks and hopes the model stitches them into truth. Access is bolted on after retrieval, provenance is a guess, and a stale chunk is indistinguishable from a current one.
Access before retrieval, evidence before answers — cited, or it says it doesn't know.
A copilot will happily draft the email, but it won't take the action under a policy, wait for the right human, or leave a record of why. The last mile — governed execution — is exactly where enterprise risk lives.
Governance before action, audit before automation.
A bigger context window holds more text in one conversation; it forgets everything the moment the session ends. Institutional memory is durable, versioned, access-controlled, and shared — the opposite of a transient prompt.
Memory is a governed system of record, not a longer prompt.
If access, citation, and audit are added after the model answers, they're theater — the model already saw what it shouldn't have. Governance has to be in the pipeline, before ranking and before generation, or it isn't governance.
Governance is a property of the architecture, not a filter at the end.
Governed enterprise memory — remembered, cited, access-controlled, and safe to act on. That's the whole product.