The four-question framework—perception, identity, memory, reasoning—is the right decomposition. Most computer vision projects stop at perception and call it done.
I trained Google Bard and have covered agentic AI extensively. What you're describing is essentially spatial-temporal reasoning, and it's genuinely difficult. The identity problem alone (is this the same physical object as before?) trips up sophisticated systems. Lighting changes, angle changes, partial occlusion—all of that can make YOLO think it's seeing a new object when it's the same one.
Your bottle demo proves the easy part works. The hard part starts now: distinguishing "this bottle moved" from "this is a different bottle" from "I'm now looking at the same bottle from a different angle."
The reasoning layer is where this gets genuinely interesting. "You forgot your charger" requires the system to understand context, not just object permanence. It needs to know a charger typically leaves with a person, recognize the pattern break, and infer intent (forgetting vs. intentionally leaving something behind).
This connects to what I'm documenting in agentic AI more broadly: memory that persists across sessions is one of the hardest unsolved problems. Most current systems are session-scoped. Building genuine continuity into physical world understanding is a much harder version of that same challenge.
Good instinct to document the build process. This kind of incremental, honest progress reporting is more valuable than a polished final demo.
