Searching your own screen

There is a kind of question that no search engine can answer: what was that thing I read on Tuesday? The page is somewhere in a billion documents, and the only index that could find it is the one nobody has built, an index of what one particular person has seen.

The published application KR 10-2026-0125947 describes a system for providing a search service over desktop and browser activity. Read through its classification codes, the design has four stages. Record what is on screen and what is heard. Extract the words, with text recognition for images and screens and speech recognition for audio. Index the result so that it can be queried. Answer questions in natural language, with a generative model composing the answer from the person’s own record rather than from the open web.

The engineering is tractable. The hard part is that such a system, by construction, holds the most complete record of a person that has ever existed. That turns every ordinary design choice into a question of trust. Should the record ever leave the device? How long should it be kept? Can anyone other than the owner query it, and can the owner tell when it has been queried? A memory that can be subpoenaed is a different object from a memory that lives in a head.

Our interest is in the version of this system that a careful person would actually accept: local by default, forgetful on purpose, and honest about what it knows.