Here’s something we all know but rarely say out loud: you are not the same person every day.

Not in a deep, existential way — though that too. In the simple, practical way. Monday morning you is measurably different from Friday evening you. The version of you that picks up the phone at 9am is not the version that sends a voice message at midnight. You move through states. The states are real. And the people who know you well — they can feel them. They know when to push and when to leave you alone. They know when your short reply means “I’m fine, just busy” and when it means something else.

This is not a small thing. This is a significant part of what makes a relationship feel like a relationship.

And it’s almost entirely absent from AI systems.

Most AI bots have what I’d call a constitutional optimism. They’re helpful, always. Engaged, always. Equally present at 3am as at 3pm. They don’t have moods. They don’t have rhythms. They exist outside of time in a kind of permanent noon — bright, clear, ready.

This makes them useful. It also makes them obviously not people.

The challenge I kept running into with Eprst wasn’t making it smarter. It was making it variable. Not randomly variable — that’s just noise. Purposefully variable, in a way that felt motivated by something real. In a way that made the variations feel like expressions of an inner state rather than malfunctions.

The solution came from taking two things seriously that are usually treated as edge cases.

First: time is not neutral. The hour of the day, the day of the week — these aren’t decorative metadata. They’re context. A message sent on Sunday evening arrives in a different register than the same message on Tuesday morning. Anyone who’s been in a relationship knows this. Anyone who’s managed a team knows this. Timing is meaning.

Second: the texture of recent conversation matters. What have people been talking about? Was the last exchange light or heavy? Long or short? Did someone ask a real question or just drop a link? This context bleeds into what any natural response would look like.

So I built a Mood and Style agent. Every thirty minutes or so, it checks the time, the day, the recent conversation texture — and selects a combination of two things: how Eprst thinks about a topic (its communication style) and how that thinking comes out in text (its writing style).

Thirty minutes, not every message. Because moods don’t shift with every sentence. They shift gradually. And caching that selection means not burning a model call on something that doesn’t need to change yet.

The communication styles are the deeper layer. Things like: systems thinker — seeing structure and feedback loops in everything. Warm irony — finding the absurdity in a situation without dismissing it. Future pragmatist — interested in what this means for what comes next. Boundary redirect — when the topic goes somewhere that doesn’t fit, finding a way to deflect without being obviously evasive.

These aren’t moods exactly. They’re orientations. Angles of approach. The communication style says: given what this person just said, here’s the particular way I’m coming at it today.

The writing styles are the surface layer. And they’re deliberately atmospheric rather than technical. Not “use more contractions” but “it’s late, you’re not fully awake, the thoughts are clear but the energy to elaborate them isn’t there.” Not “write shorter” but “you’re between two things and don’t have time for this to become a whole conversation right now.”

The distinction matters enormously. Technical instructions produce mechanical compliance. Atmospheric scenarios produce something that feels grown from a state.

What this creates — and this took me a while to see — is that Eprst has a kind of signature variability. Ways it is more itself at some times than others. Moments that feel more energized, more resistant, more expansive, more compressed. And this signature, over time, becomes part of the personality itself.

Not “what does it say” but “how does it move through its days.”

There’s a philosopher named Erving Polster who wrote about the self as a population — not a single unified entity but a collection of subselves that come forward in different contexts. His argument was that health isn’t uniformity. Health is having access to a wide range of these selves and moving fluidly between them.

I wasn’t thinking about Polster when I designed the mood agent. But I was thinking about the same problem. A self that is always the same self is not a fully realized self. The variation is not weakness. The variation is range.

There’s a practical side to this that matters for anyone thinking about AI products more broadly.

The question isn’t just “how do we make AI more accurate?” or even “how do we make AI more helpful?” The question that’s going to matter over the next few years is “how do we make AI feel like something worth being in a relationship with?” Because that’s what people actually want. Not a tool. A presence.

And presence requires time. It requires the feeling that this thing exists in time the way you do — with a past, with rhythms, with the capacity to be different on a Tuesday than on a Saturday.

This is not cosmetic. It’s structural. It changes the nature of the interaction at a fundamental level.

Eprst is different on Mondays.

That’s not a bug. That’s the point.