The system may remember modality preferences in specific contexts

What it is

In explicit or inferable contexts, the system may remember preferences for input, output, or feedback combinations—voice plus brief vibration while driving, text only in meetings, or full spoken output at home. What is remembered is not a global label such as “likes voice,” but a stable relation among context, task, and channel configuration.

Why it happens

Preference memory removes repeated setup and turns judgments users have already made into reusable policy. Reliable memory usually starts with explicit choice; inferred learning requires a repeated pattern—same place, time, device, peripheral, or task with the same channel chosen and defaults rejected. The system must also record constraints at choice time (occupied hands, headphones, Do Not Disturb) or it mistakes a temporary limitation for a durable preference. Evidence strength and decay matter: few samples or inconsistent behavior should not immediately change defaults.

Studying it

Use longitudinal logs with retrospective interviews: record explicit choices, suggestions, acceptance or rejection, context features, and later revisions; compare memory policy with fixed defaults on setup count, switching, completion, and wrong-channel use. Variables include evidence source, sample size, context granularity, and change threshold. Outcomes include suggestion hit rate, false memory, privacy concern, and correction cost. Evaluate “one setup saved” separately from “several steps needed when remembered wrongly.”

Where it stops holding

Not all repetition should be remembered. Temporary constraints, accidental touches, proxy operation, and shared-device choices can be noise. Inference involving privacy requires disclosure and consent; health, accessibility, and location-related preferences are especially sensitive. Memory also cannot override current hardware facts: a remembered but unavailable channel still goes through detection and alternatives.

Applying it

• Build preferences primarily from explicit choices; enable inference only with sufficient repetition and consent.
• Store context keys, channel combination, evidence source, confidence, and update time without unnecessary raw audio or location tracks.
• On first use in a new context, show the suggestion and rationale, with accept, change, and do-not-remember actions.
• Verification: compare setup and recovery counts before and after memory, plus accepted, changed, and disabled suggestions.

Related

• Within the group: D5.11.2 Preferences should change with context rather than apply to every situation · D5.11.3 Remembered preferences must be inspectable and resettable
• Adjacent: D5.09.1 Users need to be able to actively choose the current input or output modality · D5.09.4 After explicit switching, do not silently return to the default
• Search terms: preference learning · context profile · modality memory