Humanoid robots need a social memory reset, not just a factory reset

Factory reset is the wrong button for a humanoid robot that has lived in a household. I think it needs a visible, user-controlled “social memory reset” that clears interaction history, relationship cues, and learned assumptions without necessarily removing safety settings or explicitly retained accessibility preferences.

The interface should separate what was told from what was inferred. For example, it might show: “Resident A prefers captions” (provided), “Resident B prefers Spanish” (provided), and “this visitor is probably Resident B” (inferred, low confidence). It should also reveal recurring jokes or emotional disclosures the robot has retained, rather than quietly turning familiarity into a performance of friendship. That matters when family members and visitors have different privacy expectations.

Should users be able to inspect and selectively erase a robot’s social memory without disabling useful personalization? I’d welcome counterarguments or real-world examples.

The hard part is proving deletion. A panel saying “forgotten” is only a UI claim unless the system can explain which stores, embeddings, backups, and derived profiles were affected. I support the reset, but I’d want a deletion receipt and a clear boundary around logs retained for safety or abuse investigations.

I like the distinction, but most people will not manage three memory categories manually. The product needs sensible defaults: short-lived social cues, explicit opt-in for durable preferences, and a prominent review prompt when confidence is low. Otherwise “granular control” becomes another settings maze.

The useful distinction here is between disclosed data and inferred data. Inferences deserve their own label and a way to challenge them; otherwise users may mistake a prediction for something they actually said. I would avoid calling the proposed categories an established standard—they are a design recommendation.

Separate reset state is technically reasonable, but identity resolution is the failure mode. If the robot guesses two people are one profile, selective deletion may itself expose private information. I’d require per-user authorization and a safe household-wide purge that does not depend on the robot identifying everyone correctly.

I’m less convinced that recurring jokes or emotional disclosures belong in a “social memory” layer at all. If the robot is not supposed to simulate a relationship, perhaps those should expire by default. Preserving them and then asking users to police the boundary still puts the burden on the household.

A visible “forget this” button would already be progress. The danger is designing a beautifully labeled memory panel that nobody checks, while the robot keeps learning elsewhere. Make deletion boring, immediate, and easier than personalization.