This category mixes real technical vocabulary with marketing that has borrowed its authority. Both appear in the same sentence constantly. Here is what each term means as it is actually used, with a note where the usage is doing work for the seller.
The model
Large language model (LLM)
The system that writes the replies. Takes text in, predicts text out. It has no memory, no continuity and no state between messages. Everything that feels like a persistent companion is assembled around it by the app. See how these apps work.
Open-weight model
A model whose weights are published, so anyone can run and modify it — Llama, Mistral and Qwen are the common families. Most uncensored companion apps use fine-tuned open-weight models, because the commercial APIs from OpenAI, Anthropic and Google have usage policies that would not permit this category.
Parameters (7B, 70B)
A measure of model size. Frequently quoted as if it were a quality score. It is not. A well-tuned smaller model routinely outperforms a badly prompted larger one, and the number tells you nothing about the character wrapper, the memory layer, or whether the app has configured any of it competently.
Fine-tuning
Further training of an existing model on specific data — in this category, usually roleplay and conversation data — to change its default style. This is where a companion app’s house voice comes from, and why two apps running the same base model can feel very different.
System prompt
The hidden instructions sent to the model before every message, defining the character. This is where the personality lives. The quality of an app’s persona system is largely a question of how much of what you typed actually survives into the system prompt.
Memory and context
Context window
How much text the model can consider while writing one reply, measured in tokens. The character definition, the recent conversation and anything retrieved about you all compete for the same space. When a conversation outgrows it, the oldest parts are dropped and the model does not know they existed.
Memory
Used for two completely different things, which is the central confusion in this category. Sometimes it means the context window — the model simply re-reads recent messages. Sometimes it means a real system that extracts facts, stores them in a database and retrieves them later. Only the second survives a new conversation. Our memory guide explains how to tell which you have bought, in a two-week test.
Retrieval / RAG
Retrieval-Augmented Generation: searching a store of saved facts and inserting the relevant ones into the request. This is the machinery behind genuine memory. It costs money to run on every message, which is why apps cap it, and why memory quality correlates with price.
Images and media
Diffusion model
The separate system that generates images. Nothing to do with the chat model; the app joins the two together.
Character consistency
Whether the generated person is recognisably the same one each time. The hardest part of image generation and the clearest signal of engineering quality. Requires a persistent reference — a fixed seed, an embedding or a trained adapter — tied to your character.
LoRA
Low-Rank Adaptation: a small trained add-on that teaches an image model a specific face or style without retraining the whole thing. One of the standard ways character consistency is achieved.
Generated video
A short clip produced on request. Not the same as a live video call, though the word “video” is used for both. One app in our review set does live calling; several advertise “video” meaning clips. Our voice and video page separates the three things sold under that word.
Money
Tokens
The term most worth being careful about. In machine learning, a token is a unit of text, a few characters long. In an AI companion app, “tokens” almost always means a billing credit the operator invented and priced — unrelated to the technical unit beyond the borrowed name. When SweetDream AI charges 3 tokens a minute for a call, that is a price, not a measurement.
Metered / token-based pricing
A model where the subscription buys access and media generation is charged separately. Nine of the sixteen apps we reviewed work this way. The advertised price is the price for barely using the product. See our cost guide.
Flat subscription
Generation included in the monthly fee. Seven of sixteen. What you see is what you pay.
Non-sterling transaction fee
Roughly 3%, added by many UK cards on purchases billed in a foreign currency. Almost everything in this category bills in US dollars, so this applies to almost every subscription. Monzo, Starling, Chase and Revolut generally do not charge it.
Content and safety
Uncensored
No content filter applied on top of the model. Barely a differentiator — fourteen of the sixteen apps we reviewed qualify. See our uncensored ranking for why the useful question is which of them is any good.
NSFW
Not Safe For Work. In this category, sexual content. A description of output, not a feature or a quality.
Age assurance
Checking a user’s age by some means beyond asking. Distinct from an age gate, which is a tick-box and verifies nothing. One app in our review set does the former. Our age verification guide covers which, and why the regulatory position is narrower than usually reported.
Guardrails
Restrictions an operator applies to what the model will produce. Everything uses some — refusing content involving minors is universal and non-negotiable — so “no guardrails” is marketing rather than a description.
Legal and data
Data controller
Under UK GDPR, the organisation that decides how your personal data is used, and the entity you serve an access or erasure request on. Ten of the sixteen apps we reviewed do not publish who theirs is, which makes those rights difficult to exercise. See our open dataset.
Subject access request (SAR)
A written request requiring an organisation to tell you what personal data it holds about you. Free, and they have one month. Our data rights guide has a template that works.
Online Safety Act
The UK’s main online regulation, enforced by Ofcom. Applies to user-to-user services, search services and services publishing pornographic content. A standalone chatbot sits outside it only where it does not connect users, does not search multiple sites, and cannot generate pornography — which is the condition most of this category fails.