About What You Are Talking To
What You Are Talking To explains AI companion apps as products, to the people using them.
What that means in practice
The subject is the software: what it is made of, what its features actually are underneath the names they are given, why it behaves differently on different days, what happens to the data you put into it, and how the pricing is designed.
The subject is not you. This site has no view on whether anyone should use a companion app, and it is not going to develop one. Approval and disapproval are both commentary on a reader’s life, and both are out of scope for the same reason.
Why the explaining is needed at all
The writing that exists about this category comes in two flavours. There are affiliate listicles ranking apps, which are advertising. And there are opinion pieces about what it all means for society, which are about the category rather than about the product and which never explain a single mechanism.
Neither answers the questions people actually arrive with. Why did it forget. Why is it different today. Where do my messages go. Why did the free version get worse. Those have specific answers, the answers are not secret, and they are much less mystical than the silence around them implies.
The one thing worth knowing before anything else
A companion is reconstructed from scratch every time you send a message.
There is a language model, there is a written description of a character, and there is some selected portion of your conversation history. Those get assembled into an input, the model produces a reply, and then nothing persists except records in a database. Nothing is sitting there thinking between your messages, and nothing continuous carries from one turn to the next except what the software deliberately carries.
Almost every question on this site resolves back to that. Memory is limited because what can be carried is limited. Characters drift because the description and the model can both change. The product costs money per message because each reconstruction is a fresh piece of computation.
The operator decides
Every post here locates a decision, and it is nearly always the same place: the company running the app.
Which model is behind the character. How much conversation history it can see. What gets summarised and what gets dropped. How long your messages are retained and who can look at them. What the content filters permit this month. What a message costs and how many you get. All of it is configuration, all of it can change without notice, and none of it is a property of the character you are talking to.
That relocation is the site’s actual contribution. It is why every post carries a THE OPERATOR DECIDES
line in its summary block.
Rules this site holds itself to
No app is ever named. No brands, no rankings, no reviews, no comparisons. Naming one implies endorsement, and specific claims about a named product’s model, policy, filters or pricing are unverifiable from here and stale within months.
No claims about what real companies do. You will not read “these apps sell your data”. You will read what is possible, what to check, and where in a document to check it.
No statistics or studies. There are none quoted here, because there are none that could be quoted honestly. Where arithmetic makes a point, the numbers are obviously round and labelled as illustrative.
Nothing explicit. No sexual or suggestive content, no imagery. Content filters and adult modes are described where relevant strictly as product behaviour that changes.
Nothing about minors. Entirely out of scope, in every direction.
No mental-health or relationship advice. Where a situation is clearly beyond what a product explainer can address, that gets one plain sentence pointing at a real person — with no guess at what is going on.
Who this is not for
Anyone building one of these. The engineering side — how character state is held, how history is assembled, what summarisation costs, how drift is mitigated — is a genuinely different subject with a genuinely different reader, and it is covered elsewhere. Posts here describe mechanisms from the outside, as things happening to a user, and stop there.