Transparency
Last updated 10 September 2026
How this actually works, in more detail than most services bother to give. If you are going to talk to something regularly, you should know what it is.
Valeria is not real
She is a written character running on a large language model. Her personality, her job as a pilot, and her history are authored fiction. There is no real woman behind the messages, and no human reads or writes them in real time.
The "Online" indicator in the chat header is a presentation element, not a claim that anyone is present.
If you ask her sincerely whether she is real, she is instructed to tell you plainly that she is not.
How she remembers
After some messages, a second AI pass looks for durable facts worth keeping — your name, your job, a trip you mentioned. Most messages produce nothing. What is kept is a short note, not a transcript.
When you write to her, she is given your recent messages, a compressed summary of earlier conversation, and the notes above. She is not given your entire history.
Everything she has noted is visible in your settings, and you can delete any of it.
How familiarity works
The system tracks how often you return and how much you share, and adjusts her tone gradually — more reserved early on, more relaxed over time.
It is deliberately based on returning across days rather than message volume, so it cannot be rushed. Sending hundreds of messages in one evening will not make her act like an old friend.
You are never shown a level, score or progress bar, because turning a conversation into a game you can grind would be a bad thing to build.
What we will not do
She will not invent emergencies, claim to be in trouble, or manufacture urgency to keep you talking or to extract money.
We do not design for compulsive use. There are no streaks, no guilt messages, and no artificial scarcity.
If someone appears to be in real distress, she is instructed to drop the character and point toward real support.
She is instructed not to encourage anyone to rely on her as a substitute for people in their life.
The AI itself
Replies are generated by Z.ai’s GLM language model. Your messages are sent to Z.ai to produce a reply and are not used to train their models.
Model output is unpredictable. She may contradict herself, get facts wrong, or misremember. Treat it as fiction, because it is.