← Library · Why it's better · Test tasks

How Retrace works

Google Photos helps you search your photo timeline. Retrace helps when the timeline itself is wrong.

The problem

You remember a photo by what was happening ("when I was sick", "Riya's wedding") and roughly when. But photos copied from an old phone or saved from WhatsApp can end up filed under the day they arrived rather than when they were taken. So you search the right year and the photo isn't there, and you can't tell whether your memory or the app is wrong.

Three clocks: three evidence families

Instead of trusting one stored date, Retrace checks three evidence families and looks for where they agree.

Your lifeMoments found from your photos: trips, a move, an illness, festivals.

Your Google accountCalendar events, Gmail bookings and receipts, where you lived. Opt-in.

The file itselfThe date in a camera file name, a date printed on a paper, look-alike photos.

What you see

Every search answers "when?" "Your 'haldi' photos are from 25 Nov 2023, day 2 of Riya's wedding."
Results are grouped by moments in your life, and photos filed under the wrong date appear in the moment they really belong to: "Arrived later, belongs here".
Who is right about the date: you, or the library. "You remembered 2019. The Goa trip was Dec 2020."
Put them back: one tap moves misfiled photos to their real date, so even plain search finds them afterwards.
While you scroll, the library points out photos that probably belong somewhere else.

Why this is different

Current photo search is primarily designed to search the timeline your library already has. Retrace is designed to question that timeline when other evidence suggests it may be wrong, and to repair it. See the same memories run through both.

Questions people ask

Is my memory wrong, or the library? Either can be. Retrace shows all three side by side (what you remember, what your library says, what the evidence suggests) and how confident it is. It never assumes you are wrong.

What happens when I press "Put it back"? Only the photo's date in your timeline changes, so it moves to where it belongs. The photo itself is not edited. Undo is one tap.

Will it ever change a date on its own? No. It suggests, shows the reasons, and waits for you.

What if there isn't enough evidence? It says so and doesn't invent a date. It offers its best guess, or asks for the smallest clue you have: a moment, a year, a month, or one sentence. A month or year stays marked as approximate.

Does it read all my email? No. Calendar and Gmail are optional. If you connect them, it only looks at event titles and dates, and at bookings, tickets and receipts, near the period in question. It works with your photos alone.

Is this connected to my Google Photos? No. This is a prototype with a fictional person's library. Only the words you type are sent to the AI model; no photos or emails.

Image disclosure

The library is fictional and was made only to test this prototype. The fictional people (Aanya, her family and friends) were created from character reference portraits licensed from Adobe Stock; the scenes of their fictional life were then generated with ChatGPT image generation. Everyday photos without people are free CC0-licensed stock images (sources listed here), and the documents, tickets, receipts and chats were drawn in code. No real person's biography or personal photo library is represented.

Prototype and technical notes

Where AI is used: a language model (Claude Sonnet 5) turns what you type into a structured request: which moment, before or after it, what is in the photo, who, where, and what it was not. Finding, ranking and explaining photos is ordinary code, so the same request always gives the same answer. If the model is unavailable, a simpler keyword matcher takes over.
The library is fictional: one person, 370 items, 43 of which arrived late (copied or forwarded); 31 carry evidence that their date is wrong. Calendar, Gmail and Maps are demo data written for her.
Image understanding is simulated: what the app "recognises" in each picture (labels, people, text) is pre-written, standing in for the image understanding Google Photos already runs. "Looks like photos you took that day" stands in for image-similarity matching.
Normal search in the test is a stand-in for keyword search: it matches those labels, people, text in images, places (only when the file kept its location) and the year or month of the stored date.
19 life moments are computed by code from the photos and the account data when the library is built.