What should an AI vision assistant remember?
Thinking out loud about the difference between memory, context, and surveillance in camera-based assistants.
A camera-based assistant that forgets everything is exhausting. Every question starts from scratch. A camera-based assistant that remembers everything is unsettling, and often unnecessary.
The middle ground is short-term context. What did I just ask about? What am I still standing near? What did the last three seconds show? This is closer to how a helpful person keeps track of a conversation.
Long-term memory raises harder questions. Who owns it? Where does it live? What happens when a person wants to leave? These questions belong at the beginning of a design, not at the end.
For now, the working principle is: remember enough to be useful in the next thirty seconds. Anything longer needs a very good reason and a very clear opt in.
Tip: most browsers and screen readers can read this article aloud. On desktop, try your browser's built-in read-aloud feature.