Proactive AI companions are usually evaluated on the messages they send, but an equally important decision is when to remain silent. A companion that has many possible triggers—memories, reminders, inactivity, relationship events—can become intrusive if each subsystem sends independently. Suppression models decide which candidate messages should not be delivered at all.

Treat silence as an explicit action

The system should compare “send this message” with “send nothing” rather than assuming every valid trigger deserves delivery.

This makes restraint measurable and prevents engagement objectives from dominating user experience by default.

Use recent interaction as context

A user who just finished a long conversation may not benefit from another check-in thirty minutes later, even if an inactivity or memory trigger technically fires.

Suppression should consider time since last interaction and the intensity of that interaction, not only time since the last notification.

Learn from ignored messages carefully

Repeated non-response can indicate notification fatigue, but one missed message may simply reflect a busy day. Suppression should respond gradually rather than turning silence into a permanent disengagement label.

Explicit mute and quiet-hour settings remain stronger than inferred fatigue and should override model estimates deterministically.

Rank competing triggers before delivery

A confirmed appointment reminder, creator event update and casual relationship check-in may all become eligible at once. The system should choose the most valuable one rather than sending all three.

Trigger competition is a major source of accidental spam when product teams build proactive features independently.

Use memory confidence as a suppression signal

Weakly inferred memories should rarely initiate contact. A proactive message creates more social pressure than a normal reply, so the confidence threshold should be higher.

Confirmed events and explicit user goals can justify lower suppression, while inferred emotional states deserve much stronger caution.

Evaluate long-term outcomes, not open rate alone

A message can achieve a high open rate because it feels emotionally urgent while still increasing notification disables or churn over time.

Proactive message appropriateness benchmarks measure individual messages. Suppression models extend the evaluation to the candidate messages that should never leave the system.

A mature proactive companion is not the one that finds the most reasons to speak first. It is the one that can distinguish a genuinely useful moment from a merely available trigger and confidently choose silence when the relationship benefits more from space.

A useful evaluation compares the suppression model with a simple fixed-frequency baseline. If the learned system sends fewer messages but maintains or improves meaningful response, retention and user satisfaction, it is creating value through restraint rather than volume. Researchers should also inspect the reasons for suppression: quiet hours, recent conversation, repeated non-response, weak memory confidence or competing higher-priority triggers. Explainable suppression makes it easier to detect when the model is becoming too silent or suppressing the wrong category of outreach.

Include no-message decisions in the evaluation set

Most notification datasets contain messages that were sent and clicks that followed. A suppression model also needs cases where silence is the correct action: the user is within quiet hours, several previous messages were ignored, a remembered event has expired, or two triggers compete for the same attention window.

Compare a candidate message against a deliberate “send nothing” action. Otherwise optimization rewards finding a message for every possible event, even when the user would prefer no interruption.

Assess relationship quality after suppression

Open rate alone cannot tell whether proactive communication is welcome. Measure opt-outs, muted notifications, conversation continuation and user feedback alongside successful reminders. A model that sends fewer but more relevant messages may be better than one that drives extra taps at the cost of trust.

Long absence should not automatically trigger emotional pressure. A returning user needs an easy way to resume the relationship, not a backlog of guilt-inducing check-ins.

Suppression decisions should be explainable internally. A system may stay silent because the user explicitly disabled check-ins, because two triggers competed, or because the evidence supporting a memory is stale. These reasons have different operational consequences and should not be hidden inside one undifferentiated score.

A good offline dataset includes users who enjoy frequent contact and users who prefer long quiet intervals. Calibration should preserve those differences instead of optimizing the entire audience toward a single global send frequency. The absence of a notification can be a successful outcome when it respects a user’s boundaries.