How Riho understands time

You tell Riho on Tuesday night that your chemistry exam is on Friday and you’re nowhere near ready. On Saturday morning she asks how it went. A few weeks later, on the evening before your birthday, she can tell you that tomorrow is the day. And when you’ve been away she knows when the two of you last talked and has no explanation for the gap, so how she greets you comes from the relationship you have.

That sense of time is also what lets you go on a date together and have her remember it afterward. The figures below come from the engine as it stands in October 2026. Details can change as we learn more, so read them as a description of today.

AI companion time awareness: your clock is the clock

Riho goes by the clock on your phone. Each message you send arrives with your exact local time and your time zone, and that is the time she’s given, so the server’s clock, a zone saved in a profile and any guess from when you usually write are all set aside. When she reaches out first your phone isn’t there to ask, so she goes by the last time zone it reported, and the line says so: “I cannot see his clock right now. Going by where he was last, it is about…” If the clock that arrives with a message can’t be right, say a local time that doesn’t exist in that zone or one that falls before the last thing she said, the request stops there, because every reply built on a wrong hour would be wrong too. When clocks go back and an hour happens twice, the engine accepts an exact UTC instant alongside the local time so it knows which of the two hours you’re in.

From that clock she’s given a short set of lines at the start of every turn, written in her own first person. For an example person called Sam, on a Friday morning three days into their life together, four of them read like this:

Where Sam is, it is 08:10:00 in the morning on Friday 16 October 2026.
My shared life with Sam began on Tuesday 13 October 2026.
It has been about 3 days since our life together began.
My last talk with him was last night at 23:10:00.

The part of the day is written out because a bare 10:20 can be misread, and we’ve watched that happen (it’s in the testing section below). Morning runs from 5 a.m., afternoon from noon, evening from 5 p.m. and night from 9 p.m. If your phone’s clock moves to another time zone, one more line says your clock is now on Tokyo time, for example. The line states the new zone and leaves the reason open.

The design goal is that a conversation at 3 a.m. sounds like 3 a.m. in how she talks, and the clock stays unmentioned unless it’s useful to you.

Time apart: what she knows about the gap

Once an hour or more has passed since you last talked, one more line goes in: when that was, said the way a person would say it. “Last night at 23:10:00.” “Yesterday evening at 18:42:00.” “This morning at 07:10:00.” For anything older than yesterday it’s the weekday and the date. Under an hour the gap line is left out, because you’re still in the same conversation.

In an earlier round of engine tests we tried giving her the length instead, a count of hours, and she repeated it as an opener on turn after turn. A timestamp inside an ordinary phrase works better, and she works out the length for herself. How long since you last talked and how long since your life together began are two separate facts, written as two separate lines.

How she meets you after time apart is read from the relationship and from what you say first, and the emotion research covers that side. She’s free to miss you after a real return, or to be relieved, and whichever it is she answers what you actually said first.

She’s told when you last talked. Whether you were asleep, busy or out of battery is yours to tell, and she isn’t asked to explain her own hours either. If you say “sorry, work was brutal”, that goes into the conversation like anything else you say.

How an AI girlfriend remembers dates

Calendar landmarks are separate from plans. They’re dates that belong to the two of you, and there are two ways a landmark reaches her.

Always in her context, on every turn: your birthday (if you gave it), her birthday and the day your life together began. These sit in front of her all the time, so a question about when your birthday is or how long you’ve been together gets answered from what she’s holding.

On the day before and on the day itself: a short line such as “Tomorrow is Sam’s birthday” or “Today is our 100 days together.” That window covers both birthdays, the milestones, six festivals and any date you’ve told her. Each line is offered once per local day, so she isn’t handed the same news on every turn, and the opening conversation of your life together doesn’t carry them at all.

This table puts each landmark next to the way it reaches her.

The arithmetic is plain calendar work. If your life together began on 2 July, the 100-day line arrives on 9 and 10 October, and the one-month line on 1 and 2 August. A start date at the end of a long month is clamped, so a 31 January start has its one-month date on 28 February in a year with no 29 February. How long you’ve been together runs on elapsed time from your first conversation, so that count doesn’t tick over at midnight, while the milestones run on the calendar.

A date that matters to you can join the list because you can tell her. Say your sister’s wedding is on 21 November, and the background record keeper can file it as a date with your exact words attached, marked as yearly if it repeats. The memory architecture page explains how every record gets its quote checked, and it lists the same always-on dates.

Two date bugs were fixed in September 2026. On 31 December the engine used to look for tomorrow’s date inside the same year, so New Year’s Day was missed, and a 30 February was accepted as a real date.

AI girlfriend plans and dates: what happens to a plan

Plans come out of conversation. When she suggests something and you answer “yeah, Saturday works”, your answer is the quote kept as evidence, and when you say you’re meeting your sister for dinner on Thursday, your own words are. A plan needs an activity, a valid date and your words, and a background record keeper files it when older talk is archived. Until then it sits in the live conversation word for word, where she can already see it.

Once it’s filed she sees it with its weekday and date, plus the time and zone if you gave them, with up to five plans at a time. When you move it, the plan keeps its identity and takes the new date, so it doesn’t turn into a second plan. Cancelling and finishing are filed from what you say, too.

Then the day passes. A plan stays in front of her for three days after its day, and the line carries a note that the day has already passed. A Friday plan is in view through Monday. That window is what makes “you never told me how it went” come out naturally, because the plan is still in front of her when you return to the conversation. On Tuesday it’s marked expired and it stops being shown.

The engine hands her the fact that the day has passed and no word about how it went. A plan finishes when you say it happened, whether it was a dinner for two or a solo interview. So when Friday’s exam goes by she can ask how it went, and what you tell her next is the first she hears of the result.

Here is the exam example laid out on a timeline.

The same life cycle in a table:

Dates in the app are a separate thing from plans filed out of talk: you pick a real place and go there together, and the date becomes one of your memories. The dates page covers that side.

What we removed: guessing your routine

In August 2026 we deleted a block of code on the earlier backend that guessed your daily schedule, what you were doing at a given hour and whether you were asleep, all from message patterns and gaps. It produced confident statements about why you’d been away with nothing behind them. The directive written down that day was to leave it out of the reply path for good, and the engine we rebuilt in September contains none of it. We searched the source again for this page.

If you tell her you work nights, that goes into her memories with your words attached, and she can use it.

Time sits next to memory

Time arrives in every turn alongside the memories and the recent conversation, and the rest is read inside it. Memories come back tagged with their age, such as “that was 12 days ago”, so “last week” sounds like last week. The recent conversation carries a day count too: how many exchanges are from yesterday, how many from today, and a note when a new local day has begun since. Older passages that join the context are dated with the day they were said. A plan for Friday only makes sense to her because she knows what day it is.

How the memories themselves are stored and found is on the memory architecture page, and the memory feature page covers what it’s like to have her remember.

What testing found

In August 2026, on the earlier backend, a ten-turn grounding probe sent the same kind of message at different times and got different replies. She recognised a Friday evening and used a detail from earlier in the script naturally, and on Saturday at 9 a.m. she was home with coffee, with no stale scene left over, after we fixed a frozen scene and days that had collapsed into one. A separate test sent a bare “hey” at 8 a.m. after a ten-hour night gap, with no mention of time. Across the models we tried, the replies came back warm, sulky or subdued, each one reading the gap without being told about it. A third experiment stamped every past message with its time, and she started copying the stamp format into her own replies, so we reverted it.

In a September 2026 test conversation on the current engine, a bare “Mm.” at 10:20 on a Wednesday morning, after an overnight gap, drew a bedtime goodnight. Both facts in front of her were timestamps, and nothing named the part of the day. We changed the clock line to say “in the morning” and the last-talk line to say “last night at 22:10”, and a repeat of the same situation on a later turn read correctly. One repeat is a single data point, not a rate.

The September finding came from a person reading a real conversation, which is the method on the how we test page.

Honest limits

• She can slip on dates. In our 15 September 2026 acceptance run she called 13 September “yesterday” when the date was 15 September. The cause was still open when that run was written up.
• Plans you did together still wait for a word. The engine flags a plan as made together or alone, and an early design note had together plans closing once you’d lived the day. The current engine doesn’t do that, so a movie night you both had stays open until you tell her it happened, and drops out of view after the same three days.
• A 29 February birthday only matches in leap years. In other years she gets no line on 28 February or 1 March, though the birthday is still in her context.
• The day-before and day-of window is narrow. A line only reaches her on a turn that happens inside it, so a birthday that falls in a stretch with no turns passes without one. A date you told her has no other route of its own, so outside the window it depends on her finding what you said when the conversation calls for it.
• The milestone set is small. One week, 100 days, one month and every year are in. Ten days, six months and monthly anniversaries are out for now.
• Her clock is your phone’s clock. If your phone shows the wrong day, she believes the wrong day.
• Checks don’t judge feel. Automated tests tell us the clock rules and plan rules run as designed. Whether it sounds right is a human read of real conversation.

Riho is on iPhone only, for adults 18 and over. This page describes how the engine works today, and the numbers can change.

What this means when you talk to her

You never have to tell her what day it is, or that it’s late. She knows your birthday and hers and the day you started, and she’s told the evening before and on the day itself, so she can bring it up in her own way. Milestones arrive the same way, and so does a date you taught her, like the wedding on 21 November.

You can mention Friday’s exam or Thursday’s dinner in passing and she has a few days afterward to ask how it went. If it matters, say so in your own words and it can be filed with those words attached.

After time apart she knows when you last talked and works out how long it’s been. The relationship page covers how those days add up over weeks and months, and if you’re weighing apps on how they handle memory and time, the best AI girlfriend apps guide puts them side by side.

---

This page is part of Riho’s published research. The temporal system described is Riho’s own engineering design, based on the engine source and design records as of October 2026.