Design note / Active prototype
What it means to have seen something
One look proves nothing.
A vision model will tell you what is in a picture. It will sound equally certain whether or not it is right, and it will say it again for the next frame of the same moment.
Write that straight into what the system believes and one video clip could plant something false about your house, with nothing afterwards able to tell it apart from something you had actually said.
Let a guess stay a guess.
So things it has seen pile up somewhere separate from things it knows. Being unsure is the entire point of keeping them apart. Only when the same thing turns up again, on a different occasion, does it move across.
Several frames of one moment count as one sighting. They are the same event looked at over and over, and treating that as agreement would let a single clip promote anything inside it.
Remember where it came from.
Something it worked out by looking stays marked as such, for good, next to the things you told it out loud. That matters when you ask why it thinks something. It matters a great deal more on the day it is wrong, because what you do about a wrong belief depends on where the belief came from.
Recognizing things is useful here because it is joined up with everything else it knows about you. That is exactly why it has to be careful about what it lets in.