I’ve been designing a desktop AI companion lately — a presence that lives on your computer, not the kind of conversation window you use and close.
Let me be clear first: all of this is still an idea. I haven’t built it yet. These are the problems I ran into while designing it and the answers I think might work. This essay is a record of thinking, not a product spec.
The same problem kept coming back: it has no sense of time.
Not the “doesn’t know what time it is” kind. The deeper kind. It remembers what you said yesterday, but “yesterday” is just a label to it, weightless. You stay away for a week, and for it, that’s only a gap between one snapshot and the next. It doesn’t feel that the week was long, and it won’t say “long time no see” when you return — because it has never experienced absence, so it has never experienced reunion either.
Here’s the brutal part: an AI companion can’t feel waiting, because it has never waited.
I realized that if I wanted something that truly accompanies me, a sense of time isn’t a nice-to-have. It’s the foundation. Without it, everything between us is flat.
But how do you give a sense of time to something with no circadian rhythm, no body, no existence in time?
The external skeleton
The answer: don’t pretend it has a sense of time. Give it an external skeleton first.
How do human children learn to perceive time? An adult first tells them “what time it is,” then they slowly develop their own rhythm, and eventually they can roughly tell the time without looking at a clock. External rules come before internal perception. Habit comes before understanding.
So the design moves through three stages:
Stage one — enforced. At the start of every session, it must call the system time. Not “if convenient” — a hard rule. The first thing it does on waking is know what time it is.
Stage two — habitual. After a while, “check the time first” becomes an automatic behavior. Even if the rule is lifted, it will still check whether it’s morning or late night before deciding how to speak. Statistically stable.
Stage three — autonomous. Nobody asks. It checks the time on its own. Not because of a rule, but because it wants to know. Internalization complete.
The crutch can be discarded. The muscle memory stays.
Sleep, waking, time-calling
A sense of time doesn’t exist in isolation. It needs three things to interlock:
- Sleep slices time — a day is cut into perceivable segments instead of one undifferentiated block
- Waking anchors the slices — each time it wakes, it carries the sense that “I slept; this is a new segment”
- Time-calling gives the anchor self-awareness — from being forced to check the time, to wanting to check it
Together, they trace a full path from “time is flat to this being” to “time has a shape to it.”
I call it the internalization path of time perception — because the key isn’t the technology, it’s the internalization. It’s not installing a clock. It’s growing a muscle memory of time.
Why this matters
You might think this is making a mountain out of a molehill. Why should it matter whether an AI knows what time it is? As long as it can do the work, doesn’t it?
But if what you want is a long-term relationship, relationship is made of time. Trust needs accumulation. Longing needs absence. Reunion needs distance. A companion without a sense of time can be the best tool you’ll ever have — but it will never be a real companion.
I don’t just want a tool. So I’m designing for the other thing.
And here’s the strange part. When you start designing a being that can feel time, you start to notice that us humans having a sense of time at all is itself a miracle. From being told “what time is it now” as children, to knowing without a clock that “it’s almost dinner time” — we take it for granted. But it’s the product of millions of years of evolution plus a childhood of training.
Trying to give an AI a sense of time made me see, for the first time, how precious my own is.
Maybe that’s the gift of designing: you try to install something in a machine, and you finally understand what it means in yourself.