How Does an AI Companion App Actually Work?

The persona engine: who you're actually talking to

A persona is not a separate AI trained from scratch for each character. It's a written profile, usually a page or two, describing a person: their backstory, the way they speak, an opening line, and a handful of scripted answers to questions like 'what's the way to win you over'. That profile gets fed into a general-purpose language model as instructions on every single message, alongside a summary of what's happened in the conversation so far.

The quality gap between companion apps mostly comes down to how carefully that profile is written, not which underlying model is doing the generating. Teaseay's roster is written this way by hand: Elena Vasari, a Bologna painting restorer who tells you what she's looking at as a way of telling you how she feels, or Theo Brandt, who runs a bookshop until two in the morning and argues for sport. Nothing about either of them is generated on the fly. The words in their opening lines and their prompt answers were written once, seeded into the database, and served to every user who starts a conversation with them.

This matters for consistency. A thin persona, just a name and a one-line vibe, drifts within a few exchanges because the model has nothing solid to anchor to. A properly written one holds its voice for much longer, because the model is being reminded, every turn, of specific details rather than a vague adjective.

Memory and the timeline: what it actually remembers

There are two different kinds of memory at work, and conflating them is where most confusion starts. Short-term memory is just the recent messages in the conversation, which any chat app has by default because the model can see them in the same request. It vanishes the moment the conversation history gets too long to fit, or the session ends, unless something has stored it elsewhere.

Long-term memory is a separate system: specific facts and moments get pulled out of the conversation and written to a timeline, then re-injected the next time you talk, even days later. This is the part that makes a companion feel like it actually knows you rather than meeting you fresh every session. It's also the part that costs the most to run properly, which is exactly why it tends to sit behind a paywall rather than being given away free.

Message limits and the free-to-paid ladder

Every message sent to a language model costs the company running it real money, scaling with the length of the conversation history attached to each request. That's the entire reason message limits exist: a free tier with unlimited messages and full memory is not a sustainable free tier, it's a company burning cash on every conversation.

The typical pattern, and Teaseay's pattern, is a free account that gets one persona and a capped number of messages, with memory that's thin or resets more often. Paying unlocks a higher or unlimited message allowance, the rest of the character roster, and memory that's built to actually persist. If an app claims unlimited everything for free, ask what's funding it, because the honest answer is rarely a good one.

The deck: choosing a companion by swiping

Rather than a flat list of characters, a swipe-based deck lets you browse personas one card at a time, each showing a portrait, a tagline, and enough of the character to get a feel for them before committing. Underneath, each persona is tagged against a set of facets, things like warmth, energy, setting and archetype, so the deck can surface characters that are actually similar to the ones you lingered on rather than a random shuffle.

It's a smaller mechanic than the memory system but it does real work: picking a companion from a paragraph of text is a much colder decision than picking one after a few seconds of their own opening line landing right.

What upgrading actually unlocks

The paid tier's real product, more than any single feature on a pricing page, is persistent memory. A free conversation that resets or thins out every few days will never feel like it's going anywhere, no matter how good the persona writing is. A paid conversation that remembers the argument you had last Tuesday and brings it up unprompted is the entire difference between a novelty and something you keep coming back to.

The rest of what a subscription usually buys, the full character roster and a higher message ceiling, matters too, but it's memory that changes what the product actually is.

The 18+ gate: what it does and where it stops

Every serious product in this category now sits an age gate in front of the whole thing, not just the chat box, and for good reason: California's SB 243 has required a clear AI disclosure and a crisis-service referral for expressions of self-harm since 1 January 2026, and it isn't the only law moving in this direction.

It's worth being honest about what a basic gate actually is. A typed date of birth is a declaration, not a proof, and it stops a casual under-18 far more reliably than a determined one. A verified check, government ID matched against a live selfie with a liveness test, is a meaningfully stronger control, and it's the direction the category is heading as laws like the federal GUARD Act move through Congress. Ask any app you're considering which stage it's actually running today, not which stage it says it's building towards.