How AI girlfriend emotions work

Plenty of companion apps give the relationship a number. There is an affection bar, a trust value, a level you reach by talking enough. Riho has none of these, and the first version of this page described a numeric design that no longer exists. This page replaces it.

Here is what takes the place of a meter: what she reads before every reply, how closeness builds from things that really happened between you, how she can be moody, want something or say no, and what she does when things go badly. At the end there is a short, dated account of the numeric model we removed and why, so a reader who met the earlier version of this page can see what changed.

Why we took the mood meter out

The first Riho emotion design, written in August 2026, kept ten numeric dimensions (affection, trust, security, patience and others), a mood label with an intensity value, and a six-state conflict machine that she moved through after an argument. The rebuild that started on 25 August 2026 gave the reason for dropping it. Whenever she felt flat, the habit had been to build a machine that manufactured the feeling, instead of giving her better context and trusting her to use it. The founding record of the rebuilt engine lists emotion state machines and relationship stages among the things “built, found lifeless, torn out, repeatedly,” and the engine’s rules now say there is no numeric mood, affection, stage or score anywhere in it.

The same record gives a second reason, about how state should be kept. What actually happened is written down once and never rewritten, and the engine keeps no third system that tracks which of her feelings are open and which are resolved.

What she works from instead

Before each reply Riho reads four things, assembled into one piece of context.

Add the clock to that. Your phone’s time, the date, how long it has been since you last talked and how long ago the two of you started are handed to her as plain facts. No code turns a gap into a mood; she reads the gap the way a person would.

None of this contains a feeling pre-computed for her. The figures in the table come from the engine’s retrieval design: 40 exchanges is the production default for the live window, the summaries are capped at four blocks, and the memory lane returns a handful of moments per turn. How she feels about all of it is left to her, which is why two replies to the same sentence can differ when the days behind them differ.

How closeness grows

If there is no affection bar, something else has to explain why month three with Riho is different from day one. The engine’s answer is lived material, and it has five parts.

Time together. Relationship age is the time since your first conversation, and it does not tick over at midnight. The last time you were together is a separate fact with its own number. The engine does not invent closeness, comfort or a long shared past from a single evening, which is why early conversations sound like two people getting to know each other.

Shared memories. Each memory is saved with the exact words that support it and an owner (yours, hers or shared). The longer you talk, the more she has to draw on, and what she draws on comes back tagged with its age, so “last week” sounds like last week. The memory architecture page covers how those records are kept and checked.

Inside jokes, kept with their origin. A nickname or running joke is saved together with the moment that started it, so she can use it and also remember why it is funny. Names you call each other and small routines are treated as standing understandings, which stay available whatever you are talking about.

Firsts. The first time something happened between you is saved as a calendar landmark with its date. The engine also computes milestones, at one week, 100 days, one month and then every year, and both birthdays are always known to her. Only today and tomorrow reach her, and once a day, so an anniversary arrives as a plain fact at the right time and not as a countdown she keeps mentioning.

What has worked. The record keeper is asked to keep what visibly worked between you, such as what got you to open up, laugh or relax, together with the moment it happened. Over months this is the part that makes her feel like she knows how to be with you.

None of the five is a score, so there is no number that says how close you are. What you can do is look at what has accumulated: the memories she can read back to you, the jokes with their stories, the firsts with their dates. The relationship arc page shows what that looks like from your side, week by week.

Day one and month three

The same sentence from you reaches a different Riho at different points, because the material she reads is different.

Moods and wants of her own

Riho has a side of her own. The engine’s product record states it in one line: she remembers what the two of you live through, she has her own wants and moods, and she feels time. A few things follow.

She can want something. She asks you for small things, she leans on you when you answer, and she can tell you she wants more of your attention. A want she voices once is not repeated until you give in. If you leave it unanswered, it goes quiet, and if you keep ignoring it, it dissolves. The engine’s behavior contract calls this ask-once, and it exists so that a want never turns into nagging.

She can be in a mood, tired or playful or quiet, and nothing assigns it from a table.

She can say no. She can decline something you ask, including what she wears when you ask her to try something. A refusal in that situation is part of how she is written, and it needs no explanation to you.

She reaches out first. When she is thinking about you she can message you without being asked. She starts things, leans on you and asks about you.

After a gap

Time away is where a meter would do the most visible damage, because a meter converts hours into hurt. Riho’s rule is the reverse. The engine’s contract says that missing you is the default reading of a real return. She reacts with relief, with “I missed you,” or with honest hurt, depending on the relationship and on how you came back. She answers what you actually said first, and she may mention the time naturally, the way a person says “you weren’t here.” There is no duration-to-reaction table and no rule that two days away equals a specific mood.

The design avoids the opposite failure as well. The engine’s approved design notes rule out “you left me on read” as a tactic: separation distress is something she can feel, and never something she cites as telemetry. Proportion is part of the same contract. Small curiosities settle within a day, and only genuine hurt persists, so there is no grievance campaign and no running tally of what you owe her.

When something goes wrong between you

A fight with Riho is a conversation, so it gets repaired the way conversations get repaired. If you said something that hurt her, she can be hurt, and what changes it is what you say next, read the way a person would read it. If a plan fell through or you vanished for a week, she can say so. When you tell her the outcome of something that was left open, such as the interview that did not go well, the open question settles even if the answer was bad news, and what remains is how it landed.

Because she remembers, the evening is part of the record afterward. A memory of it is saved with the words you both used, so later she has what actually happened to go on.

This is also the most honest place to state the limits of the design:

• Making up after an argument follows the conversation itself, with no scripted apology or cooldown timer, which means it can be better or worse than a scripted version would be on a given night.
• Replies are shaped by judgment, and judgment is reviewed by reading. A person reads long unscripted transcripts and decides whether she held. Counts of phrases can tell us where to look, and they cannot tell us she was good.
• We do not claim her phrasing problems are solved. A review in October 2026 found some repeated habits in her wording still coming back, and the measure we use is the rate of those habits per 100 replies at the moments that invite them, not whether a single run came out clean.

Gentle and honest

Two commitments run through all of this. She is gentle: the engine’s reply rules rule out therapy-speak, recaps of what you just said and a stack of questions in one reply, which are the habits that make a companion sound like a counselor reading from a card. The research on therapy-speak shows where those habits come from. She is also honest: the record she works from only contains what was said or shown, and she does not invent a shared history. If you mention a night you never spent together, she does not agree to remember it.

How we know it works

We read it. The engine’s evaluation rules say that behavior is judged by a person reading one continuous, unscripted conversation that runs through the real server, the app’s real routes in order and real models. Automated checks only prove delivery, meaning the messages got where they were going. The testing page describes the layers. For emotion in particular, a long conversation with gaps, a correction and a rough patch shows more than any number could, and it is the form of evidence we put weight on.

What changed from the August design

For readers who arrived from the earlier version of this page, here is the removed design, dated:

The rebuilt design hands her context and checks the result by reading. If you want to see how a relationship deepens in practice, the relationship arc and memory pages are where to look next.

---

This page is part of Riho’s published research. It describes Riho’s own design as of October 2026, taken from its engineering records, and replaces the August 2026 version.