Daily Living
Cognitive Companion also watches for the rhythm of an ordinary day: has she eaten, how long was the TV on, does today's outfit look like yesterday's, does it look like she's about to make tea. These signals feed two things: a countable record a caregiver can ask about later, and a small number of gentle, evidence-backed alerts. None of it replaces a caregiver's judgment; it gives them something to act on with confidence instead of a guess.
Ask the companion about the day
Once activity detection rules are configured for a household, a caregiver can ask the voice companion questions like:
- "Has she had her medication today?"
- "How many times did she eat today?"
- "How long did she watch TV this afternoon?"
These answers come from a structured daily activity record, the same kind of record a caregiver would keep on paper: an entry opens when the activity starts and closes when it ends, with a room, a time, and a note on how confident the system is. The companion never guesses a count by skimming its memory of what it saw; it always reads from this record, so "three times" means three recorded meals, not three vague impressions.
Every record also carries where it came from. A meal or TV session detected from cameras and sensors is phrased differently from a medication dose confirmed step by step during a guided routine with the resident: the companion says "she completed the medication routine with me at 9:05" only when that is what actually happened, and something more hedged, like "she was near the medicine cabinet, but I can't confirm she took them," when the evidence is weaker. A false "yes, she took her medication" is a real risk for a dementia resident's care team, so the companion is built to be honest about how sure it is, not just confident-sounding.
Hygiene: same clothes, and a missed wash-up
Two things feed into a single hygiene signal for the caregiver:
- Appearance comparison. Each day, the system builds a summary of what she was seen wearing, drawn from her clearest camera sightings that day, and compares it to yesterday's summary. A day-over-day match that is unusually close is a candidate for "may be wearing the same clothes as yesterday."
- A wash-up check. Separately, the system checks whether she spent a plausible amount of time in the bathroom in the last day or so, the kind of dwell that suggests a shower or wash-up happened.
When both point the same way, an image comparison model looks at her best photo from each of the two days side by side and confirms whether they genuinely look like the same outfit before anything reaches the caregiver. Only when the clothes plausibly match and there was no plausible wash-up does the caregiver get a "possible missed hygiene routine" alert, with the two comparison photos attached as evidence.
Privacy note. Attaching the comparison photos to the alert is the default, since seeing the evidence is part of what makes the alert trustworthy and actionable. A household can turn this off to send a text-only alert instead, keeping the underlying detection the same.
Where this stands today. The day-to-day appearance comparison has been tested against weeks of historical data to calibrate how close a match needs to be before it is worth a look; live alerts stay off until that calibration is judged solid enough with real production data. Until then, this is a built and tested capability, not yet a live one.
Noticing tea-making, gently
The caregiver mentioned that the resident won't ask for help and that her tea time is unpredictable, so a fixed reminder schedule does not fit. Instead, the companion watches for the pattern of getting ready to make tea: lingering in the kitchen, near the kettle. When it looks likely, the caregiver gets a notification with a snapshot, nothing more.
The companion does not yet start the guided tea routine on its own from this signal. Before that happens, the caregiver reviews each of these notifications and marks whether it was actually right (she was making tea) or not, building up a track record. Only once that track record consistently shows the detector is accurate does the household turn on the next step, where the companion offers to walk her through the tea routine step by step when it notices the same pattern. That is a deliberate settings change, not something that turns on by itself, and it is additionally held back until the guided companion's own safety and escalation systems are fully in place: a wrong, unprompted conversation with a vulnerable person is a real cost, so this one earns its way to full autonomy with evidence, not a code change.
How this stays affordable
All of this analysis runs on the household's own hardware, shared with everything else the companion does. To keep it responsive, every one of these checks is designed to reuse information already computed wherever possible (where she is, how long she's been there) before reaching for a more expensive look, and a heavier "describe what the camera sees" or "ask an AI model about this image" step only runs when a cheaper check has already made it worth doing. Camera analysis is also skipped automatically when a scene plainly has not changed since the last look. The heaviest checks are also capped so only one runs at a time on the shared hardware, so a burst of activity around the house is handled in order instead of overwhelming the system.