Most AI waits. The useful part is going first.

Nearly every AI assistant is pull-based. You open a window, type, and get an answer. Close the window and the relationship pauses until you restart it.

That model is fine for questions and poor for commitments, because commitments fail in the gap between conversations — precisely where a pull-based tool has no presence.

Sending an unprompted message is easy. Sending a useful one is not. Ari has to know what you committed to, when it is due, what you already did, and what you said weeks ago that makes now the right moment.

That is why memory and a real calendar are prerequisites rather than features. Without them, proactive messaging is just notification spam with better grammar.

An assistant that messages constantly gets muted, and a muted assistant is worse than none. Every unprompted message spends trust.

So the useful design goal is not more messages. It is fewer messages that each land at a moment where you would have wanted the nudge.