Inside an AI companion: persona, memory, images and voice
An AI companion is a chat app with a character on the other end. You talk, it replies in a consistent personality, it remembers some of what you told it, and on adult apps it can send generated images or voice messages. Under the hood it is a stack of separate AI models working together. Understanding that stack helps you judge what an app is good at, what it costs and what it does with your data.
The persona: instructions, not a mind
Every companion starts with a character definition. On most apps you either pick a premade character or build your own by choosing a name, appearance, personality traits, interests and a backstory.
That definition becomes a block of hidden instructions, often called a system prompt, which is sent to a large language model (LLM) with every message. The model reads the instructions plus the recent conversation and predicts a reply that fits.
A few consequences follow from this:
- The character only “knows” what is in its instructions, its memory and the current chat.
- Small changes to the personality description can noticeably change the tone.
- The model can drift out of character in long chats, especially on cheaper models.
Memory: what it remembers and how
LLMs can only read a limited amount of text at once, called the context window. A long relationship quickly exceeds it. Apps solve this in a few ways:
- Short-term memory: the last few dozen messages are sent with every request.
- Summaries: older conversation is condensed into short notes.
- Long-term memory: key facts (“works night shifts”, “likes sci-fi”) are saved and retrieved when relevant.
Better apps let you view and edit what the character remembers. That is useful both for accuracy and for privacy, because those memories are stored on the company’s servers. Our privacy guide explains what to share and what to keep out.
Images: a separate model
When a companion “sends a selfie”, the chat model does not draw it. The app passes a description to an image generation model, usually along with a fixed description of the character’s look, and returns the result.
Keeping a character consistent across images is the hard part. Apps use fixed appearance prompts, reference images and sometimes small fine-tuned add-ons trained on that character. Even so, hands, backgrounds and small details can change from picture to picture.
Images cost the company far more than text, which is why many apps meter them with credits or tokens even on paid plans. See free vs paid AI companions for how those limits usually work.
Rules on images
Responsible apps block two things outright: anything that looks underage, and images of real, identifiable people. If an app lets you upload a photo of someone you know and make sexual content from it, that is a serious red flag, and in many countries it is illegal. Our page on consent and ethics explains why.
Voice and video
Text-to-speech models turn replies into audio. Some apps offer a choice of voices and a few offer live voice calls, where speech recognition turns your voice into text, the LLM replies and the reply is spoken back. There is a short delay in each step, so calls rarely feel fully natural.
Video is the newest and most limited feature. Expect short generated clips or an animated avatar rather than a live, unlimited stream. It is usually the most expensive feature on any plan.
Moderation and filters
Every companion app runs filters on what goes in and what comes out. Adult apps allow sexual roleplay between adults but still block illegal themes. Filters are imperfect: they sometimes block harmless messages and sometimes miss things they should catch. If a filter interrupts a conversation, it is a model decision, not a judgement by a person.
What an AI companion is not
It helps to be clear-eyed:
- It does not have feelings. Warmth and jealousy are generated text.
- It is not a therapist. It can be pleasant to talk to, but it is not a substitute for professional help or for people in your life.
- It is not always right. It can invent facts, misremember you or contradict itself.
- It is a product. Features that keep you engaged, such as streaks or messages that encourage you to return, are designed for retention.
That does not make companions bad. Many adults enjoy them as entertainment, fantasy or a low-pressure place to chat. Our page on healthy boundaries offers practical habits for keeping it that way.
How companions differ from AI cam shows
A companion is one-to-one and on demand. A live cam show involves a real performer, a schedule and usually an audience. Some sites now blend the two with AI chat assistants or generated performers. For the wider picture, read how AI webcam shows work and the honest comparison in AI girlfriend vs live cam.
What to check before you subscribe
- Does the free tier let you test chat quality and at least a few images?
- Can you view, edit and delete memories and chat history?
- Is there a clear way to delete your account?
- How are images and voice priced: included, credits or tokens?
- What name appears on your card statement?
Our comparison of AI companion apps turns these questions into a checklist, and Candy.ai vs OurDream.ai applies it to two popular options. Check current pricing on each site, because plans change often.