Almost every complaint about these apps — the repetition, the drift, the uncanny inconsistency between two runs of the same scene — comes from one mechanism. It is worth ten minutes because it turns "the app is broken" into "the app is doing the thing it does, and here is the setting."
It writes one piece at a time
The model does not compose a reply and then type it. It produces one token — roughly a word or part of a word — then reads everything including that token and produces the next.
That is the whole loop. There is no plan, no outline, no draft being revised. A reply that ends beautifully did not know it would when it started.
Two consequences follow immediately, and both are things you have seen:
A reply cannot un-commit. Once the model has written "I have never been to Paris," everything after is conditioned on that. It will build consistency around a sentence it produced by accident rather than contradict itself.
Errors compound forwards. A slightly-off word in sentence one nudges sentence two, which nudges sentence three. This is why long replies drift and short ones rarely do.
The dice roll between each word
At every step the model has a ranked list of candidates with probabilities:
maybe "good" at 30%, "fine" at 12%, "terrible" at 3%, and a long tail.
If it always took the top candidate, the character would be deterministic and extremely boring — the same greeting every time, the same three jokes. So the app samples instead: it rolls weighted dice.
The weighting is controlled by a setting usually called temperature. Low temperature concentrates the probability on the obvious candidates: consistent, predictable, eventually dull. High temperature flattens the distribution: more surprising, more creative, and more likely to produce something that does not follow.
You rarely get a slider for this. What you get is an app's chosen point on that trade-off, and it is a large part of what people mean when they say one app "feels smarter" than another. It is not always smarter. Sometimes it is just warmer or colder.
This is also the honest answer to "why did regenerating fix it?" Nothing was fixed. You drew again from the same distribution and got a better roll.
What the model is actually looking at
Before your message, the app assembles a block of text you never see: the character description, the facts it has stored about you, a summary of your history, and the recent conversation. That whole assembly is the context window, and the reply is generated from it as a single continuous document.
This has a practical implication most people never exploit: the model responds to the shape of what it can see. Give it three short, flat messages and it will produce short, flat replies, because that is the pattern it is continuing. Give it a vivid message and the register shifts. You are not persuading a person, you are setting a pattern, and the pattern is contagious in both directions.
It also explains the most common self-inflicted problem in this category. People settle into "hey", "how was your day", "what are you doing" — and then conclude the app got worse. The app is continuing the document you are writing together.
Why apps sound different when the model is the same
Several apps in this category run on similar underlying models. They still feel distinct, and the differences come from things layered around the model:
- The system prompt — how the character is described, at what length, with what instructions about tone and pacing.
- The sampling settings — the temperature point discussed above.
- What gets retrieved — which facts and which slice of history are put in front of the model for this particular turn.
- The filter — what gets refused, and whether a refusal is graceful or a wall.
Nomi reads as consistent over weeks because of the third item, not because of a bigger model. Candy AI covers chat, images and voice in one subscription, which is a product decision rather than a modelling one. Neither difference would show up in a benchmark, and both show up within a fortnight of use, which is how we score them.
Four things this lets you do
Steer early, not late. The first two sentences of a reply determine the rest. If a scene is going wrong, stop it and restate, rather than arguing with paragraph four.
Use regenerate deliberately. If a reply is 80% right, regenerating throws away the 80%. If it is wrong in its first sentence, regenerate immediately.
Write the register you want back. Short and flat gets short and flat. This is the single cheapest improvement available to anyone who thinks their companion has become boring.
Do not argue about facts it invented. A model that has written something false will defend it, because consistency with its own output is what it is built to do. Correcting it in a new message works; demanding it admit the error does not. That behaviour has its own article: why AI companions invent things.
The part worth remembering
There is no one home between your messages. The character exists for the duration of a single generation and is rebuilt from text at the start of the next one. Every impression of continuity is an achievement of the plumbing around the model — the stored facts, the summaries, the retrieval — and that plumbing is what actually differs between a $10 app and a $20 one.
Which is why our ranking weighs memory and consistency as heavily as it does. They are the parts a landing page cannot show you.
