The AI That Knows You Longer: Building Cleo

The problem I kept running into

Most AI companions are impressive for about a week. They can hold a conversation, sound warm, even feel perceptive in the moment — and then you realize they have no idea who you were yesterday, let alone who you were three months ago. Every conversation starts from zero. That's fine for a assistant. It's a dealbreaker for anything trying to feel like a relationship.

The current wave of emotionally expressive voice AI is very good at reading this moment — tone, energy, what you're feeling right now. What almost none of it does is notice change — that you sound different than you did last month, that something in how you're showing up has shifted. That noticing is what makes another person feel like they actually know you. It's also the single hardest thing to fake, and the thing no one has really built.

That's the problem behind Cleo: an AI companion built around getting to know someone better over time, not just responding well in the moment.

The core idea

Cleo is built on a simple bet: the relationship should compound. Every conversation should leave Cleo with a slightly better understanding of who you are, not just a longer transcript to search through later. Over time, that accumulated understanding is what lets her notice something a good friend would notice — that you seem off, that something's changed, that you're not quite yourself — and gently name it, rather than treating every conversation as a fresh start.

That capability doesn't come from a bigger model or a better voice. It comes from a genuinely different kind of memory — one that tracks who you are and how you're trending over time, rather than just storing what was said. It's the difference between a friend who remembers facts about you and a friend who notices you've changed.

Why this matters

A few convictions shape the direction of this project:

• Recognition isn't understanding. Reading someone's mood in real time is valuable, but it's not the same as knowing them. Understanding requires memory that deepens, not just responsiveness that's sharp in the moment.

• The relationship should get harder to replace, not easier. A companion that's interchangeable with any other chatbot the day you start using it was never really a companion. The value should compound the longer you stay.

• Trust has to be actively protected. A system that's building a longitudinal model of someone carries real responsibility — the person needs control over what's remembered, visibility into what's being noticed, and the ability to have some conversations not count at all.

• This is a relationship, not an assistant. Cleo doesn't take actions on your behalf, doesn't manage your calendar, doesn't send messages for you. Her job is to know you — not to do things for you.

Where this is headed

The current build proves the core idea end-to-end: real voice conversation, a memory that builds a picture of the user over weeks rather than resetting every session, and the ability to notice and gently name meaningful change. From here, the focus is validating that experience with real users beyond just internal testing, and thinking carefully about the ethical guardrails a product like this requires as it scales — informed consent around what's being remembered, and making sure the relationship stays healthy rather than something someone becomes unhealthily dependent on.

I'm always glad to talk through the thinking in more depth — just reach out.