The persona goes in first
Before your message, the model receives the character definition: who she is, what she does, how she talks, how she relates to you. This is why a companion with an authored persona stays in voice and a generic one drifts into sounding like an assistant. The model is not remembering to be her, it is being told who she is every single turn.
Then the conversation, then the memory
Recent messages go in verbatim, because that is what makes replies coherent turn to turn. Older history cannot: a context window is finite, and a year of conversation does not fit in one. So platforms summarise and store the durable parts, then reinject the relevant ones.
That is the actual engineering problem behind AI companions, and it is where they differ most. Good memory feels like she knows you. Bad memory feels like she is doing an impression of someone who read your file once.
How the photos stay her
Image generation on its own produces a different woman every time you press the button, because a prompt is not an identity. Platforms that keep a face consistent attach a fixed visual identity to the character record and apply it to every render, so the prompt controls the scene and the character controls the face.
On HoneyPot that identity is locked when the companion is created, which is why a gallery of thirty renders reads as one person and why a wizard-created companion looks the same six months later.
Voice, and why it is billed differently
A voice call is a live pipeline: speech to text, model, text to speech, in a loop tight enough to feel like talking. It costs real compute per second of conversation, which is why calls are billed per minute rather than folded into a flat allowance.