How do AI girlfriends work, under the hood?

Three separate systems, chained together: a language model that writes the reply, a memory layer that stores what you've told her, and a voice model that turns the reply into audio when you call. None of them is doing all three jobs at once.

Does she actually remember previous conversations, or is that faked?

The memory is a real store outside the model — names, dates, preferences — pulled back into the conversation when it's relevant. It's not the model re-reading the whole chat history every time.

Why does voice take longer to respond than text?

Text streams as the model writes it, so you see it almost immediately. Voice has to wait for the full reply, then run it through the voice model before you hear anything, which adds real seconds.

What decides how explicit a reply gets?

The Lust Level slider — five steps, set by you — controls what the model is allowed to generate. It's a parameter passed to the model, not a mood it reaches on its own.

Are the photos pre-made or generated live?

Generated live, from an image model working off the character's established look. That's why the same request can come back slightly different each time.

18+

Three separate systems write every reply

How do AI girlfriends work? Most apps hide the pipeline — this page names every layer instead of guessing.

Flirt now Watch her

Six characters, six separate personality profiles

Each one runs on the same underlying models but a different personality profile and memory store — that's why they don't feel interchangeable after the first few messages.

Online

Voice

Live

Under the hood

What's actually running when you chat

Six things happening behind a normal-looking chat window: the model, the memory, the voice, the render, the latency, and the setting that controls all of it.

A language model writes the reply

Every message goes through a language model tuned on the character's written personality, not a static script. That's why two chats with the same character diverge depending on what you say.

70+video modes

Tap one, she does it.

A separate layer stores what you said

The model itself doesn't remember earlier sessions — a memory layer outside it logs names, dates and running jokes, then feeds the relevant lines back in before she replies.

Where the seconds actually go

Tap one, she does it.

Flirt now

Synthesized, not a recording

Her voice comes from a voice model trained on a single actor's read of the character, run on whatever reply the language model just wrote. Nothing is pre-recorded.

Photos generated on request

Ask for a photo and an image model renders one inside the thread, using the character's established look and the scene you're both in — not a fixed set you're picking from.

The Lust Level is a setting, not a mood

Five steps control what the model is allowed to write and generate, from company to explicit. It's a switch you flip, not something she arrives at on her own.

The pipeline

What happens between your message and her reply

Four steps, four separate systems. None of them is instant, and none of them is quietly doing another one's job.

1
You send a message
Text or voice, it enters the same pipeline. Voice gets transcribed first; typed text skips that step entirely.
2
The model generates a reply
The language model writes the reply using the character's personality profile, the current Lust Level, and whatever the memory layer just fed it.
3
Voice or photo gets rendered
On a call, the reply gets synthesized into audio in her voice. If you asked for a photo, an image model renders one in the same beat.
4
Memory gets written back
Whatever mattered in that exchange — a name, a plan, a preference — gets written to the memory layer so the next session starts from where this one left off.

Flirt now

Against the typical app