AI companion research

Riho’s research pages are written by the people who build her, from the engine’s code, our design records and dated test runs. Ten pages are published. They cover how she remembers, how she knows what day it is, how her character and feelings are put together, and how we test any of it.

Most pages carry two kinds of statement. Some describe the design as it is in October 2026, and others report what a test found in July, August or September 2026. We label both, because the engine was rebuilt in September and an August result can describe a system that no longer exists. Where this hub and a page disagree, the page is the one to trust.

Every research page and the question it answers

AI companion memory research

Two pages cover memory, one for what you say and one for what you show her.

How her memory is built and checked

Memory architecture follows what happens between mentioning an interview on Tuesday and her asking how it went on Thursday. Everything you say is saved word for word, a background record keeper turns older talk into short records, and each record has to quote the words that support it. Code looks for the quote in the cited exchange and rejects the record when it is missing, and a separate ownership rule means a fact about you has to come from your own words. The page also lists where this breaks: corrections can miss, and in a run on 15 September 2026 she called 13 September “yesterday” on the 15th.

What she keeps when you show her a photo

Photo memory looks at what she keeps of a picture and how she tells it apart from another one that looks like it. She keeps the photo with a description of what is visible and the words you sent with it, and what the photo meant is read from the conversation around it. In a live run from 22 to 24 September 2026, a recompressed copy of a photo was matched to the original showing, and a wider shot of the same scene was treated only as resembling it. A stale record also stayed active after a correction in that run, and the repair delivered on 23 September has not been confirmed in a live run.

How she knows what day it is

She goes by the clock that arrives with each of your messages, and temporal awareness covers how that clock, the gap since you last talked, birthdays and milestones reach her. In August 2026 we deleted code that guessed your daily routine from message patterns, and the rebuilt engine has none of it. In a September 2026 test conversation a bare “Mm.” at 10:20 on a Wednesday morning drew a bedtime goodnight, until the clock line began naming the part of the day. A repeat of the same situation read correctly, and one repeat is a single data point.

Character, emotion and voice research

Four pages cover who she is and how she sounds, and each one records something we changed.

Why companions sound like therapists

Therapy-speak asks why an AI companion answers a bad day with “I’m here for you”. In July 2026 we ran five prompt versions on a small model, and banning phrases only moved the therapy-speak into new wording. What worked then was a character bible plus code that trimmed and filtered her replies. That code is gone from the current engine, and a voice review in mid-October 2026 still found some repeated habits in her wording.

How her character is written

Character design traces five iterations of writing a personality down. Numeric traits such as “humor 0.8” gave flat conversations, prose bibles of about sixty lines worked better, and roughly half of each early bible described what the character refuses to do. An audit on 5 October 2026 reversed part of that, with a new rule that nothing in her prompt should be a sentence she could say to you.

Why “never do X” held up better than “be X”

In the July and August 2026 setup, positive rules like “be snarky” faded within three to five turns, while anti-rules like “never be sentimental” held for the session. Prompt methodology is that method, and it also records the catch the October audit found: a rule written in her voice becomes a line she can say.

How her feelings work without a mood meter

Emotion and conflict explains how her feelings work with no numbers behind them. Our August 2026 design had ten numeric dimensions and a six-state conflict machine, and the 2026 rebuild removed both. She now reads the latest exchanges, short summaries of older talk, memories found by meaning and standing facts, with the clock alongside, and closeness comes from time together, shared memories, inside jokes and firsts. She can want something, be in a mood, say no, and message you first.

How we test and what the tests can show

Three pages cover testing, from how we test Riho to how you can test any companion.

How Riho is tested

How we test says a person reads long, unscripted conversations with her, run through the same server, requests and order the iPhone app uses, and that pass or fail is that person’s read. It lays out what each kind of check proves and what it can’t, from automated tests that show delivery to volume tests and a planned long blind relationship test, which has not been run. The page keeps the August 2026 instruments as labeled history: a 30-day relationship arc, a 16-case girlfriend eval, an existence test across many systems and a grounding probe.

A method you can run yourself

The memory benchmark methodology is written so another person can repeat it on any app. It defines six scenarios from fact recall to delayed recall, controls that prevent accidental hints, a run log for build, plan, device and dates, and six outcome labels from Exact to Unavailable.

What the first probes found

The August 13, 2026 probe notes are the dated results so far, on Riho only. A watch habit was stored, retrieved and placed in the reply prompt, and then one run invented that the user was handling the watch while another left it unmentioned. A moved pillow was not stored as a memory, and a plan could be created, moved and closed with no duplicates. These results predate the September rebuild, and the page assigns no score.

AI companion benchmark use case: choosing between apps

Apps describe memory in their own words, which makes them hard to compare without a test of your own. The methodology looks at four practical questions: whether a detail you chose to save comes back, whether its meaning survives, whether a correction is accepted, and whether the current version is used after a delay or a switch between text and voice. You can run a first check over two days. Share three details with each app, correct one of them, wait a day, then ask about all three in chat and again on a call. Write down the app version, the plan and each reply, and use the same script every time. The memory benchmark methodology turns that into a fuller test, and what it is like to have her remember shows the product side of the same questions. If you are still choosing, the best AI girlfriend apps guide compares them, and how Riho works puts chat, memory, photos, calls and dates side by side.

How to read these pages

Riho has not published a scored comparison of companion apps on memory, and none of these pages names a winner. The only memory results published so far are the Riho-only probes from August 13, 2026.

Voice problems are reported as a rate per 100 replies at the moments that invite them, since one clean run can’t show how often a habit comes back.

Several pages list their own failures, among them a correction that left an old record active and a date she got wrong. We leave those in because they are the part another team can check against its own system. If a page and the live app disagree, tell us through the contact page.

Riho is on iPhone only, for adults 18 and over. The design on these pages is the engine as of October 2026, and it will change as we learn more.