Myeverly

Myeverly started from an article about how large the AI-companion space had become. I wasn’t interested in the surface of it — I was interested in the hard technical problem underneath, and whether the people using these products were actually being served well.

The Thesis

The interesting problem in AI companions isn’t the personality — it’s memory. How do you store what an AI knows about a person so it persists? How do you keep that memory from drifting or being destroyed when the model underneath gets swapped for a new one? And how do you do all of that in a way that’s careful about the mental-health dimension of these products, which most builders were ignoring?

I framed Myeverly as a privacy-first companion built around a durable memory architecture — the layer that would keep “who you are” intact even as the models changed.

What I Built

A working prototype. I built out full personas and tested them, and — critically — I did the research. I reached out to prominent, heavy users of AI companions and had hours of detailed conversations about how they actually used these products, the technical limitations they hit, and the workarounds they’d invented to preserve their companion’s identity over time.

The Insight That Reframed It

My initial assumption was that users would want well-crafted, pre-built personas to choose from. The research said the opposite. The most invested users already had a companion — one they’d shaped over months — and what they wanted was to import it somewhere the memory wouldn’t drift. They didn’t want a new character. They wanted continuity.

That completely changed how I understood the product. The value wasn’t the persona; it was the persistence.

Why I Sunset It

The technical problem stayed fascinating. The market around it did not. Most of the space was NSFW or built around loneliness, there were well-funded incumbents already entrenched, and I didn’t want to build something that could contribute to mental-health harm — even a well-intentioned version of it.

So I sunset a product that worked. The memory-architecture question — how to preserve accurate, non-drifting information about a person across model generations — was too good to throw away, and it carried directly into how I think about AI systems in my later work.

What I Learned

The best user research is the kind that changes your mind. It reframed the product for me in a single conversation — and then it told me something harder to hear: that this wasn’t a space I wanted to build in. Walking away from a working build is a decision, not a failure, and being able to make that call cleanly is its own kind of product judgment.