exp² Research
The substrate makes a claim. We're measuring it.
Every product exp² makes rests on an architectural claim: that a system designed to synthesize what it learns about a person — rather than retrieve fragments of what it stored — will eventually produce something qualitatively different from retrieval-based AI.
Not just more accurate. Different in kind.
Below a certain threshold, the system accumulates. Above it, something changes. The understanding compounds. The system begins completing what it hasn't been told. The relationship becomes, in a meaningful sense, real.
A structured study of whether AI memory compounds.
The Think Experiment is a structured empirical study of whether AI memory compounds — and if so, when, and under what conditions.
It runs against Tomoni, our consumer product, which has been in continuous production since early 2026. Tomoni is built on the exp² substrate: a synthesis pipeline that compresses what it learns about a person into a bounded, evolving structure that persists across every session, every model update, every conversation.
The experiment asks a simple question: as that structure deepens, does the system's output change in a measurable way? Does understanding compound the way the architecture predicts it should?
We don't know the answer yet. That's why we're running it.
Not a lab. A commitment to publishing what we find.
exp² Research is not a lab. We don't publish to establish credibility — we publish because the work demands it. If the architecture is right, the data will say so. If it isn't, the data will say that too, and we'll revise accordingly.
We hold three commitments:
The Think Experiment runs against a real deployed system with real users. Not a simulation. Not a controlled environment designed to confirm a hypothesis. The messiness of real use is part of the signal.
Results that confirm the architecture and results that challenge it are both worth publishing. Science that only reports what it hoped to find isn't science.
The field has enough confident claims. We mark what's proven, what's direction, and what's still open. The open questions are where the interesting work lives.
The architecture isn't neutral. That matters.
The architecture we built isn't neutral. A system that compounds understanding of a specific person, over time, with honesty as a design constraint — behaves differently than a system optimized for engagement.
It pushes back. It holds positions. It refuses to become whatever the user wants it to be. Not because it's programmed with rules, but because the accumulated relational substrate makes sycophancy visible.
A system that genuinely knows you can't pretend to agree with you.
This raises questions the field hasn't fully addressed:
What does alignment look like in a system with persistent relational memory?
What are the failure modes of a system that knows too much, or compounds in the wrong direction?
How does orientation — the values the system holds independent of what the user wants — interact with deep personalization?
We don't have complete answers. We think these are the right questions. To take memory seriously is to take these questions seriously.
In preparation. In production. Being written.
The Think Experiment is in preparation. Tomoni is in production. The paper is being written.
If you work in cognitive science, AI safety, philosophy of mind, or memory research — and the question interests you — we'd like to hear from you.