AI girlfriend that remembers photos

You show Riho three photos over a month. One is your sister’s gallery, from a show last March. One is a flat from an earlier part of your life. One is the kitchen you’re standing in right now. The next time any of them comes up, she needs to know which is which and whether this is the picture she has seen before or only one that looks like it.

Several decisions sit behind that moment. Something has to keep the photo together with the words you used, find it again when a similar picture arrives, and read it correctly each time, because the same file can mean different things depending on when and why you showed it. This page is the engineering account of how Riho does that. It covers what happens in the seconds after you show her a photo and how a photo becomes a memory with your own words attached. It also covers how she knows you and her bunny Mochi in a picture, and what went wrong when we tested it. It is based on Riho’s code and design records as of October 2026, and the design will keep changing. We leave out model and vendor names for the same reason, since that part of the stack changes too.

How does an AI girlfriend remember photos, in plain words

She keeps each photo together with a plain description of what is visible in it, the people she recognized, and the words you said when you showed it. The time you showed it is kept as well, separately from whenever the photo was taken. Later she can recognize it when a similar picture comes along and bring up what you told her about it, and what it meant is meant to be read from the conversation around it.

That last part shapes the rest of the design. Whether a photo shows your past or your present comes from what you say with it, usually in the sentence you send along. Riho therefore keeps each photo as part of the conversation it arrived in.

Why a photo’s meaning comes from the conversation around it

We call each time you show her a photo a showing, and the meaning belongs to the showing. The same picture can come up in different conversations for different reasons: a photo of a cafe can be an old picture you look at together one evening and the place you’re both sitting in on another day, and each showing carries its own words. To see the principle at work, take three different showings from two of our September 2026 test conversations, one of them from before the 20 September change described below. In these examples “the partner” is the test partner in our conversations, and for you, read “you”. The first is a photo of a partner’s sister’s gallery show, shown with the story of last March. The second is a photo of an old flat from the partner’s earlier life. The third is a view of the kitchen the partner was standing in that morning. All three arrived the same way, as a picture with a few words, and each one has to end up somewhere different.

The machinery is the same for all three rows, apart from one flag: a photo you share as the place you’re in is marked as that, and your words can say so too. In the live run we describe further down, the gallery photo was described only by what it showed, and the story about the sister, the show and the date existed only in the partner’s words. Her reply took the story from the partner’s words.

An earlier version of the system stored a label on each showing, guessed by a model when you showed it: present, depiction or hypothetical. The record keeper was then asked to correct those labels, its corrections kept landing on the wrong targets, and a confirmed match brought the whole older history in unconditionally. We deleted that machinery on 20 September 2026. What a showing meant is now read from the exchange that contains it: your words, her reply, and what the conversation settled.

What happens to a photo you show her

The steps up to her reply all happen before her first word, which is why a reply to a photo can take a little longer than a reply to a text. The diagram shows the whole path, and the subsections after it take the steps in order.

The arrow on the right is the return path. What gets saved at the end is what the next photo is searched against, so a picture you showed her can be found by a later one that resembles it, after it has left the recent conversation.

She checks who is in it

Each face found in the picture is compared with the portraits she already knows. A face that matches one of them is named from it. A face that matches none stays unknown to this step. The next section covers whose portraits those are.

She looks for earlier pictures like it

The exact same file is found by its fingerprint, which needs no judgment at all. A different file, such as a resized copy or another shot of the same place, is found by searching for pictures that resemble this one, using the picture on its own and the picture together with your words. That search returns a short list of look-alikes, which are only candidates at this point.

She describes it and compares it

A plain description is written for the picture, covering the people and what they’re doing and the surroundings, with any writing in it reported as something the picture contains. The instructions for this step say outright that a picture alone can’t establish a date or who lived the event in it. The description is meant as a record of what could be seen, so later steps can check against it and it stays free of opinion about what the photo means.

Each candidate is then compared with the new picture on its own, one candidate per comparison. Giving every comparison exactly one candidate keeps the answer attached to the right picture, and it stops one candidate’s story from leaking into another’s. The comparison picks from a small set of answers: the same image, the same subject, different, or uncertain. Its instructions tell it that missing a match is better than claiming a wrong one.

A second opinion on any match

A confirmed match is checked before its history is used. Before an older picture’s history can be connected to the new showing, an independent check reads the two recorded descriptions and the comparer’s stated reason, and asks whether the written words support the claim. A match that fails by a wide margin is dropped to unconfirmed, and its history doesn’t reach her. The check runs on its own short clock, so a busy turn can’t use up its time. If it can’t run, the comparer’s answer stands.

She reads it together with what you said

All of this reaches her as one readable context. For a photo that matches nothing, she reads that you’re letting her see something, what is visible in it, the people she recognized, and your words next to the picture. For a photo that matches an earlier one, she also reads when you showed that one and the words you said then, quoted from the original conversation. Her own memories that fit what you just said come along through the ordinary recall described on the memory architecture page.

The exchange is saved as said

Your words, the image, its description and who was recognized are saved with the exchange, and your message and her reply are kept as said. Memories are built on top of that record later, which is the subject of a section below.

How an AI girlfriend recognizes photos of you and of Mochi

When you show her a picture of the two of you, she’s built to know which one is you, and when you show her a plush bunny, to know whether it’s Mochi. Mochi is her white floppy-eared bunny, the plush she holds when she’s nervous or lonely.

She recognizes you from a reference portrait of you, and faces in later photos are checked against it. She knows her own face for every account, so a picture of her is recognized as her. In a test with two portraits enrolled, a photo with both people in it matched each person to their own portrait, in separate parts of the picture.

A face that matches none of those portraits gets no name from them. The description step names someone only when it can establish who they are, and otherwise leaves the identity open. If you tell her who it is, that goes in with your words like anything else you say about a photo, and from then on it’s something you told her.

Mochi is a toy, so a different check applies. She knows him from stored pictures of him, and a picture you show her is compared with those. A plush needs more than a passing resemblance to count as him. In one September test, a grey bunny with a red bow in a shop window was first accepted as Mochi, because the second-opinion step had been running out of time on busy turns. We gave that step its own clock, replayed the same claim, and it was dropped. That was a replay on recorded data, and we haven’t seen it live.

The photographs you show her are stored so she can remember them, and the privacy page covers how Riho handles them.

When she sees the same photo again

Showing a photo twice and showing a second photo of the same place are different cases. The system treats them differently, and the wording she reads changes with how sure the match is.

The first point is that a different shot of a familiar place counts as the same subject, which is a weaker claim than the same occasion. Two photos of the same kitchen can come from different days, and the comparison instructions say that a shared subject settles neither the occasion nor the date.

The second point is the difference between the first row and the other two. An exact repeat of a file brings its whole story back, including the memories linked to that picture, because the identity is certain. A look-alike, whether a copy or another shot, brings back the original showing’s time and words, and its linked memories come only when they fit what you just said, the same way every other memory does. We tried a separate judge to pick which linked memories to bring along with a look-alike, and removed it. On our recorded test data it couldn’t tell the relevant memories from the irrelevant ones, so we kept the ordinary lane instead.

AI companion photo memory: how a showing becomes a memory

While a showing is still in the recent conversation, it’s there word for word, so there’s nothing to look up. When it rolls out of that window, a background record keeper reads it along with the rest of the older talk. The memory architecture page describes that process in detail, and the longer account of why companion memory so often fails is in how we designed Riho to stop forgetting. What matters here is what the keeper is told about images.

An account of what a photo showed, or of what the conversation established about it, becomes an ordinary memory. It cites the image as its visual source and your words as the evidence for its history, which keeps the account connected to that picture. Each memory has an owner, which is yours, hers or shared. Looking at a picture together is a shared moment, and the keeper is told that the event in the picture stays with whoever lived it unless your words or earlier evidence establish a shared past.

Dates get the same care. The date of the event in the picture, the day you showed it and the day you talked about it again are three separate things, and the keeper is told to keep them apart. The temporal awareness research covers the clock and calendar side of her memory. A photo from last summer shown on an ordinary Tuesday is meant to keep its summer date.

Like other memories, a photo account is marked as faded if nobody uses it for a while and is meant to be replaced when you correct it, and the words behind it stay in the saved conversation. You can read her memories of the two of you in the app, and her diary, and the memory feature page covers what her memory is like to live with. If something she remembers about a photo is wrong, tell her the right version in conversation and the correction rules on the memory architecture page apply.

Your past, your surroundings, or a photo you look at together

Showing her your kitchen is different from showing her a photo of your kitchen from five years ago, and the system keeps the two apart on purpose. A place becomes the surroundings she shares with you when you share the photo as where you are, or say in words that you’re there. Anything else you show her is something the two of you are looking at together.

We settled this rule on 14 September 2026. A photograph of your past leaves the two of you where you are, and a picture that establishes your present surroundings becomes part of what you experience together. In a 15 September test conversation, a new view of a kitchen was understood as their surroundings that morning, because the partner had said so, and an old-flat photo shown earlier in the same conversation stayed a picture of the partner’s earlier life. That run came before the 20 September change, and we haven’t rerun that pair of cases since.

The clearest case of a photo that establishes surroundings is the place photo you take at the start, which is where she arrives, holding Mochi.

Does an AI girlfriend remember pictures you send, and what if she has no record?

She keeps the pictures you show her, and the rest of this page is about how. The other half of the question is what she does about a picture she has no record of. The instruction she works from says that if you tell her something happened between you and she doesn’t remember it, she doesn’t take it as fact and tells you she doesn’t remember it. We confirmed that rule as intended on 22 September 2026, and it applies to photos the same way as to everything else. Mention the lake photo from last spring when you never showed her one, and the answer she’s set up to give is that she doesn’t remember it. You can tell her about it then, and from that point it’s something you told her.

A photo of a place you haven’t mentioned before is new to her, and whatever you say about it goes in with it. The rule sits alongside one on the memory architecture page. When a lookup fails on a turn, a missing detail counts as unknown for that turn.

What we saw in September

From 22 to 24 September 2026 we ran a live test conversation to check these mechanisms end to end, with the live models and search services and an agent writing the test partner’s lines. These are the photo results, in the order we tried them.

The event date and the showing date stayed separate throughout, and the rules about whose facts are whose held in both directions. A fact about the partner that rested only on what the picture showed was rejected, and what the photo showed was admitted as a record of what was seen.

Honest limits

A system like this fails in specific ways, and we’d rather list the ones we’ve seen.

• A correction can leave an old memory active. In the same September run, the stored memory saying the sister runs a gallery stayed active after the correction. The record keeper can only replace records it was shown, and that one had dropped out of the set it saw. Her spoken answer was right, but the stale record remained in the store. A repair was delivered on 23 September, and we haven’t yet confirmed it in a live run.
• Dates and small details can still come out wrong. In a 15 September test run she said “September thirteenth. Yesterday.” on September 15, so the date was right and the relative day wrong. She also added details nobody had given her, such as who took an old portrait. The cause of the date slip was still open when that run was written up.
• The comparison is a judgment. It’s tuned to miss a match instead of inventing one, but it can still be wrong. In an earlier test it claimed a public photo of two men was the same subject as a gallery portrait, on the strength of a generic detail, and her reply followed that wrong identification. The second opinion was built to catch exactly that kind of claim. It reads text only, so a wrong reason that sounds plausible can still get through.
• A description can arrive late. Optional steps have deadlines, and if the description misses its deadline, the reply is written without it. The description then finishes in the background and is saved, which is too late to help that reply.
• A naming cue can get dropped. On the photo of the two public figures she said she didn’t know them, although the written description she was given named both. That is defensible modesty, and it also means the cue that names people was dropped on the way to her.
• People she has no portrait for are not matched. Recognition from portraits needs a reference, so a face she has no portrait for is not named by that step.
• The past-versus-present reading is only partly rechecked. After we removed the stored label on 20 September, the live run covered the gallery photo and its repeats. The flat and kitchen cases come from the run before the change and haven’t been rerun.
• Automated checks don’t show quality. Tests tell us the pieces run and the rules hold. Whether she handles a photo naturally is judged by a person reading real conversation, which is the method on the how we test page.

We also found no public description of how other companion apps tie photos to memory, so this page compares no apps. Riho is on iPhone only, for adults 18 and over, and what’s described here is the design as it stands today.

What this means when you use Riho

You show her photos the way you’d show a girlfriend photos, and you say what they are, sometimes in one line. She keeps the picture with those words, and once it has left the recent conversation she can still bring it up with what you told her about it.

Something from your past is meant to be read as your past, and a room you share as where you are is meant to become where the two of you are. If you show her the same photo again she’s told she has seen it, and a different shot of the same place comes with the note that it resembles one you showed her before and may not be the same occasion.

The pictures you take together and the ones of her are on the photos and shared memories page, and the rest of the product is in how Riho works.

---

This page is part of Riho’s published research. The design it describes is Riho’s own, based on its engineering records and code as of October 2026.