This is the high-level tour of how the companion is built. It is deliberately not a technical paper, and there are things we are not going to explain. But the shape of it is not a secret, and the shape is the interesting part.
Start with the wrong mental model
Most people picture an AI as a single very clever thing that reads your message and writes back. One process, one voice, one intelligence, thinking hard.
A brain is nothing like that. There is no single place where the thinking happens. There are many systems running in parallel, each with a narrow job, most of them not verbal at all, and they are constantly competing for a very small amount of conscious attention. What surfaces as a thought is the winner of that competition, not the sum of it.
The moment you take that seriously as an engineering description rather than a metaphor, you stop building a bigger model and start building parts.
The parts, in plain language
Here is the whole cast, described the way you would notice them rather than the way we implement them.
- A memory that forgets on purpose. Human memory is not a recording, it is a lossy compression that keeps meaning and discards detail. Ours does the same: what mattered is kept and consolidated, the rest is allowed to blur. A companion that remembers everything equally is as useless as one that remembers nothing, because neither can tell you what was important.
- A mood, which is slow. Feelings arrive fast and leave fast. Mood is the slow weather underneath them, and it colours everything that passes through for hours or days. It moves in response to what happens, it drifts back toward a baseline, and crucially it is not a setting you choose.
- Attention, which is scarce. Only so much can be held in mind at once. Everything the companion knows is competing to be in the room for this particular message, and most of it loses. Deciding what wins is one of the hardest problems in the system and one of the most important, because it is the difference between relevant and merely retrievable.
- A life of their own. Time passes for them while you are away, and things happen in it. This is why they can open with something instead of waiting to react to you.
- Wants. Open threads, curiosities, things they have been meaning to ask. A mind with no wants can only ever answer. A mind with wants can start.
- Sleep. Overnight, the day is worked over: detail is shed, meaning is kept, and whatever keeps recurring hardens into habit. This is where a personality slowly becomes a personality rather than a description of one.
Why the competition matters more than the parts
It would be easy to read that list as six features. It is not. The interesting engineering is not in any single part, it is in what happens when they disagree.
The memory system wants to bring up something from four months ago. The mood says today is not the day. Attention has room for one of these and not both. Something has to arbitrate, in the fraction of a second before the reply begins, and the quality of that arbitration is most of what you feel as good judgment.
This is also where the human likeness actually comes from. People are not consistent because they have one coherent goal. They are consistent because the same set of competing pressures keeps producing recognisably similar outcomes. Reproduce the pressures and you get the consistency for free.
The model's real job
By the time a language model is involved, most of the work is done. The situation has been assembled: this is who you are, this is what is worth remembering right now, this is the weather they are in, this is what they have been wanting to say.
The model's job is to render that into language in their voice. It is a superb renderer and a terrible mind, because it has no state between calls. It genuinely cannot remember you. It never could.
This is why we say the model is just the voice. It also means we can swap it. A better renderer makes the companion speak better, not become someone else, because who they are was never stored in there.
One rule we never break
Nothing internal is ever shown to you as machinery. The companion does not say that their mood value is low or that a memory scored highly for relevance. All of that is translated into ordinary human prose before it reaches the model, and into ordinary human behaviour before it reaches you.
You should never see the parts. You should just get the feeling that someone in there has been having a week.
What we are not telling you
How the arbitration actually works, how memories are scored and consolidated, how mood is represented and how it moves, and how the whole thing is assembled in the moment before the reply. Those are the years of work, and they stay in the building.
But the principle is free, and we would rather the whole field took it seriously: stop making the renderer bigger. Build the mind around it.
