Recommendation engines need context modes, not one permanent taste profile

I want recommendation systems to distinguish “I’m researching this” from “I like this.” A temporary work project or sensitive research topic can otherwise become evidence about my permanent taste, shaping recommendations and interface behavior long after the task is over. In the worst case, those suggestions reveal what I’m working on to someone looking over my screen—or simply make the product feel uncomfortably observant.

I’d rather have explicit modes such as Work, Research, Shopping, and Leisure. Each mode should show which signals it uses, how long they persist, and whether they can affect the default profile: never, with approval, or automatically. There should also be a visible mode indicator and a straightforward reset/delete control.

This isn’t anti-personalization. It’s more accurate personalization: purpose is not the same thing as preference. Would people actually use these controls, or would the extra choice become another setting everyone ignores?

The strongest precedent is fairly modest but useful: YouTube tells users to pause watch or search history when researching something for a school project they are not personally interested in. That supports the problem, though not necessarily the full mode proposal. The missing piece is visibility: users should not have to know a hidden workaround exists.

The hard part is not adding a mode selector. It is preventing leakage between profiles. A system can display “Research” while still using shared embeddings, ranking caches, or derived interests in the default feed. The boundary needs to exist in the data and model pipeline, not just in the UI.

I like the framing because it gives the product team a better promise: “more relevant in this context,” rather than “we know you forever.” I’d make the default behavior conservative, then ask whether a user wants to promote a saved item or signal into their general profile.

A visible indicator matters more than it sounds. Without one, people will forget which context they are in and accidentally search for personal things inside a work mode—or vice versa. Also, deletion should cover derived recommendations, not just the raw click history.

I’m not convinced most people want four modes. Many will pick one and never switch, producing another false sense of control. A simpler “temporary session” with an expiry and an explicit “add this to my profile” action might capture most of the benefit.

There is a support angle here. When recommendations go strange, users currently have to guess which old activity caused it. Showing active signals and their expiry would turn a mysterious personalization failure into something support teams—and users—can actually diagnose.

The retention choice should be per mode, not a single global setting. A shopping basket may need to persist for a week; a sensitive research session may need to disappear when the tab closes. Purpose limitation is practical when the product makes the purpose explicit.

The industry has spent years making personalization invisible, then acts surprised when it feels creepy. A small label saying “based on your research activity” would at least make the bargain legible. I’d take that before another clever preference dashboard.

One implementation question: what happens when the user is logged out on one device and logged in on another? If modes sync, that creates more exposure; if they do not, the experience gets inconsistent. Still worth solving, but the threat model belongs in the first design review.