Search and insights

Search answers "get me back to that moment." Insights answers "where did the day go." Both read the same local database, and neither sends a query anywhere.

Two search modes

LUCI's search box has two modes, and they fail in different ways, which is why both exist.

Keyword matches the literal text that was on screen. Use it when you remember an exact string — an error code, a filename, a person's name, a URL. It searches the whole capture, not one line, so a query whose words landed in different parts of the screen still hits. Chinese and pinyin queries are handled by a dedicated tokenizer rather than being split on spaces.

Semantic matches meaning. Use it when you only remember the gist: "the article about database migrations", "that pricing table someone shared." It works on any language you can describe the thing in, including one different from the text on screen.

Spoken content is searched the same way. A question that spans a meeting and a screen — "what were we looking at while we talked about the migration?" — is one query, not two.

Every search takes an app filter and a time range, and results come back as moments: what was on screen, what app, what window, and when.

Enhanced Search

Semantic search has a stronger mode you turn on yourself.

With Enhanced Search off (the default), LUCI answers semantic queries by narrowing with keywords first and re-ranking those candidates by meaning. It is fast and costs nothing extra to store, but it can only rank things a keyword already found.

With it on, LUCI stores a vector for every block of text it reads, searches those directly, and backfills everything already captured so your history becomes searchable by meaning too. Better recall, more disk, more work per captured frame.

Find it in Local Settings → Storage → Enhanced Search. It is read when the app starts, so the change lands on the next launch.

Insights

The Insights dashboard lives on Settings → Account. It has two halves.

Recording. Hours actually recorded, your top five apps by time, and a drill-down per app showing its busiest window titles and its own per-day history. Screen time is measured from the real gap between frames, with a cap so a machine left asleep does not bill you eight hours.

Agents. How much context your agents pulled out of the memory, how many calls they made, and an estimate of the time that saved. There is a 26-week heatmap you can flip between recording days and agent days, a streak count for each, and a list of recent calls.

Pick a range of 7 days, 30 days, or all time.

How "time saved" is calculated

The number is deliberately conservative — it is meant to survive a skeptical reading.

The ceiling matters: one call can serve an hour of screen history, and without a cap that single call would claim hours saved, which nobody would believe. The breakdown is also in the app, behind the number itself.

Every figure on the page is computed on this machine from your own capture and call logs.

Related

• Search your memory for how to phrase a question
• Agent Bridge for the tool calls Insights counts
• Your Life folder for the written daily record