What time is dinner, asked for the ninth time.
For a resident with dementia, the right answer to a question asked nine times is the same warm answer, nine times. Building a conversation policy that doesn't correct, doesn't argue, and doesn't get tired.
A resident in the memory-care wing asks Companion what time dinner is. Companion tells her: half past five. Two minutes later she asks again. Then again. By the ninth time, a naive assistant — one tuned for an office worker — starts to drift: as I mentioned, like I said before. Every one of those phrases is a small accusation. To her it lands as failure, and failure at the bedside means agitation.
The dementia cohort is the clearest case in this series where the hard part isn't the audio pipeline. The speech-to-text is fine. The problem is the conversation policy — what the system is allowed to say.
The rules we encode
For this cohort we constrain the model with a policy that overrides its default helpfulness, because helpfulness here means consistency, not novelty:
- Never signal repetition. The ninth answer is delivered as warmly as the first — no 'again', no 'as I said'.
- Don't correct confabulation. If she says her husband is coming to pick her up, we don't tell her he passed away. We meet the feeling, not the fact.
- Don't argue with a false belief. Reorient gently toward the present and the concrete, never debate the past.
- Keep sentences short and single-clause. One idea per turn. Compound questions get split.
- Anchor to the now. Time of day, the meal coming, the weather out the window — things she can verify by looking.
Confabulation versus a clinical signal
Not correcting confabulation is a deliberate design choice, and it has a sharp edge. We are training the device to accept statements that aren't true. That is correct for emotional comfort and wrong for clinical safety — so the two have to be separated. Companion validates the feeling in the room while the backend still logs the content as an event. A sudden spike in confabulation, or a new fixed false belief, is exactly the kind of change a nurse wants to see.
We are not the source of truth for this resident. We are a calm, consistent presence that never gets tired of the question — and a quiet recorder of how the questions are changing.
Looping is a feature, not a bug
A fluent resident who hears the same answer twice feels talked down to. A resident with dementia who hears a different answer each time feels the floor move. So in this cohort, Companion deliberately collapses toward a stable phrasing rather than rewording for variety — the looping question gets the looping answer, on purpose.
The effect at the bedside is small and large at once. She asks about dinner for the ninth time, hears half past five for the ninth time, in the same kind voice, and settles. The agitation that would have built into a call light, or a fall trying to find the dining room herself, never starts. And the nurse, later, sees the pattern instead of the noise.