Inspiration

Memory loss in older adults starts with small things: a door left unlocked, a pill skipped or taken twice, a stove left on. Most memory tools expect you to remember to use them, with reminders to set, apps to open and buttons to press. That breaks down for exactly the people who need it most. We wanted something that pays attention for you and speaks up only when something was actually missed.

What it does

Memory Band runs on a Vexo wristband and works passively: the wearer never has to say or press anything.

  • Front door: if you leave home without locking it, about 20 seconds later your phone says "You didn't lock the door." When you go back and lock it, the alert closes by itself with a follow-up message.
  • Medicine: it hears the pill bottle rattle and feels your hand move to your mouth, so it knows when a dose was missed or taken twice.
  • Daily habits: it catches a toilet that wasn't flushed, too long without a drink of water, and a phone set down and left behind.
  • Safety: it notices a possible fall and gives you a countdown to say you're OK before family are alerted. It also tells you when you forgot to put the band on.
  • Dashboard: family get one large-print page that shows the problems first and everything else as a simple checklist.

How we built it

  • Band to computer: a Python bridge connects to the band over Bluetooth using the Vexo Open Kit and streams its microphone, motion sensors and taps.
  • Sound and motion detection: NumPy code picks out the start of each sound and compares its frequency pattern against saved examples. Each sound has its own rule. A flush is 2.5 to 10 seconds of steady rushing noise, and a sip is a hand-to-mouth lift of 0.3 to 4 seconds with no pill rattle nearby. The band counts as off the wrist when its motion sensors stay below their normal noise level for a while.
  • Server: FastAPI with SQLite stores events and runs a rules check every 5 seconds. Each alert goes from open to pushed to resolved, or escalates to family, and phone notifications go out through ntfy.
  • Setup: a guided page where you make each sound once while wearing the band. Any step can be skipped, and the system falls back to built-in defaults.

Challenges we ran into

  • Telling everyday sounds apart. A lock click, a pill rattle and keys set on a table can all be short, sharp sounds. Each needed its own timing and frequency rules, plus motion signals, to tell them apart reliably.
  • Knowing that something didn't happen. Most systems react to events. We had to detect missing ones, like leaving with no lock sound or a toilet-paper sound with no flush after it, without making noise over false alarms.
  • Bluetooth limits. The band accepts only one connection at a time and drops when you walk away. We had to handle reconnecting and keep the "connected" status honest. Early on the dashboard showed the band as connected and not connected at the same time.
  • Designing for older users. Our first dashboard looked like a typical tech product. We rebuilt it with large print and plain words, put problems at the top, and kept "skip and use the default" available on every setup step.
  • Testing without the real event. We couldn't flush a toilet at the venue, so we wrote 64 automated checks on synthetic sounds and motion, plus a pretend-band mode for the full flow.

Accomplishments that we're proud of

  • It's truly passive. The wearer never opens an app, says a command or presses a button. The band just notices everyday sounds and movements.
  • It catches things that didn't happen, not just things that did: a door not locked, a pill not taken, a toilet not flushed, no water for hours.
  • The setup is easy. You make each sound once while wearing the band, or skip the step and use the built-in defaults. No technical knowledge needed.
  • It ends with a fix, not just a warning. Alerts follow through to resolved and shows how long the fix took, so family see the problem was handled, not just that it happened.
  • The detection code passes automated checks, and a pretend-band mode lets us test the whole flow without hardware.

What we learned

A tiny wrist microphone hears a whole house, so sound alone is rarely enough. Most of our best results came from pairing a sound with a movement. A pill rattle followed by a hand-to-mouth lift means a dose was taken. A lock click with a wrist twist means the door was locked. We also learned to design around absence: the most useful alerts come from what should have happened and didn't, which means thinking in time windows and expected sequences, not single events. The Bluetooth connection, sensor noise and battery-saving behavior changed our assumptions several times, so we built automated checks to keep moving when the band wasn't connected.

What's next for Memory Band

  • Test detection on real recordings from many homes and bathrooms.
  • Use location to learn which sounds belong in which room.
  • Add a weekly summary for family that shows trends, not just alerts.

Built With

Share this project:

Updates

Submission history