Skip to main content

Memory

Neural Cube keeps a memory of the people it has met, the places it has been, and the things it has been told. It is what lets the robot get better at interacting with people over time, rather than starting from a blank slate every session.

Memory is fully on-device. Nothing the robot remembers is uploaded, shared, or synced to an outside service. A simplfied view of the memory:

META-COGNITIVEmornings are the busiest hoursChanel keeps returning to musicSteve prefers short answersLASTINGChanel was reading a novelChanel takes her coffee blackChanel plays the pianothe dog is called BiscuitSteve's birthday is in MarchRECENTtalked about a trip to KyotoChanel asked for calm pianoI suggested a light dinnerSteve is out until FridayPLACESthe kitchen is to the leftChanel's mug lives by the sinkthe front door is past the hall
meta-cognitive — patterns noticed across many memorieslasting — durable facts about a person, place, or topicrecent — context from the conversations of the last few daysplaces — named locations, attached to the map
learned togetherconsolidated fromrecalled for this turnfading
A memory graph mid-turn. “Chanel asked for calm piano” has just been stored, and it pulls in the lasting fact it belongs to and the patterns built on top of that (the ringed nodes).

How memories shape future interactions​

Before the robot responds or acts, it pulls in the memories relevant to what is going on right now. In practice, this shows up in small, cumulative ways:

  • Familiar people are addressed by name, and the robot adjusts to how each one prefers to interact.
  • Preferences only have to be mentioned once.
  • Places the robot has learned the name of can be referred to by name later.
  • Context about a project, a routine, or an ongoing situation is there the next time it comes up, so the robot does not have to be caught up from scratch.
  • Things the robot has been asked to keep an eye out for stay on its mind across sessions.

Older memories that stop being referenced fade in influence, so the robot's picture of its environment reflects how things are now rather than how they were a year ago.