AI Companion Glossary: Friend, Character, Agent, Memory, and More

These AI companion terms are often used as if they were interchangeable: friend, companion, character, chatbot, assistant, agent, memory, context, and more. They are not fixed scientific categories, and product marketing regularly stretches them.

This glossary gives each term a practical meaning and explains why the distinction matters when choosing an AI companion. It is a map, not a claim that every company uses the words consistently.

The relationship words

AI companion

An AI companion is a conversational AI product designed to support an ongoing, relationship-style interaction rather than only complete isolated tasks. Companions often combine a persistent identity, memory, proactive contact, and text, voice, or images.

That is the broad AI companion meaning. The category can include platonic friends, romantic companions, character-based relationships, and other forms. It does not mean the software is conscious, human, or qualified to provide care.

A current US Senate bill proposal defines an AI companion around adaptive, human-like responses and the simulation of interpersonal or emotional interaction; that legislative definition is useful evidence of the emerging category, though proposed legal language is not a universal dictionary.

AI friend

An AI friend is a companion framed around friendship rather than a romantic, professional, or purely fictional role. The product may remember personal context, keep a consistent point of view, and initiate conversation.

The term describes the intended experience, not a human mind inside the software. Read the fuller answer to what an AI friend is and the boundary-focused discussion of whether AI can be a friend.

Virtual companion

Virtual companion is an older and broader label that can include conversational software, game characters, digital pets, embodied avatars, or social robots. Not every virtual companion uses generative AI, and not every AI companion has a visual body.

Social chatbot

A social chatbot is built for open-ended interpersonal conversation rather than a narrow service flow. Microsoft researchers used this term for XiaoIce, a system designed around long conversation and social engagement. The XiaoIce research helps show how the companion category grew from social-chat research.

The product-role words

Chatbot

A chatbot is any software that communicates through a conversational interface. It may follow a fixed decision tree, retrieve approved answers, or generate open-ended language. Chatbot describes the interface more than the relationship.

An AI friend is a kind of chatbot, but most chatbots are not friends. A parcel-tracking bot can be friendly without maintaining a personal relationship. See AI friend vs. chatbot for the complete comparison.

Virtual assistant

A virtual assistant helps a person get something done: find information, set a reminder, send a message, operate an app, or complete another task. Apple describes Siri in those terms.

Assistants can have memory and personality. Their main success measure is still utility or task completion. An AI companion’s main job is the continuing conversation. The AI friend vs. virtual assistant guide explains the overlap.

AI agent

An AI agent pursues a goal through multiple steps and may use tools or act in external systems. It might search, compare, schedule, send, monitor, or retry. The meaningful difference is authority to act, not whether the language sounds more intelligent.

A companion can be proactive without being an agent. Sending “How did the presentation go?” is conversational initiative. Booking a new presentation time through a calendar integration is external action and needs a different permission model.

Virtual character or character chatbot

A virtual character is a generated persona with an identity, backstory, voice, appearance, and behavioral definition. It may exist for roleplay, interactive fiction, entertainment, tutoring, or companionship.

Character.AI’s current Creator Guide describes names, avatars, voices, greetings, tags, descriptions, and definitions as the pieces that shape a Character. A character becomes companion-like when the product and user invest in an ongoing personal relationship rather than primarily in a role or scene. Read AI companion vs. virtual character for the decision guide.

Persona

A persona is the set of instructions and facts that shape how an AI presents itself. It may include tone, habits, preferences, boundaries, backstory, and example dialogue.

A persona supports consistency. It is not evidence of consciousness. If a companion says it went to pottery class, that can be a coherent part of a disclosed fictional persona without being a literal event experienced by software.

The generation words

Generative AI

Generative AI produces new synthetic content—such as text, audio, images, or video—based on patterns learned from data and the input it receives. That broad definition aligns with the NIST generative AI glossary.

For an AI companion, generation may create the reply, a voice rendering, or a fictional photo. Generated does not automatically mean false in a harmful sense; a joke or imaginary scene is supposed to be invented. It does mean the output should not be presented as verified human fact.

Large language model (LLM)

A large language model is a machine-learning model trained to process and generate language. It can write fluent replies by predicting tokens in context. An LLM is often the conversational engine inside a companion, but it is not the complete product.

Identity, stored memory, timing, delivery, billing, moderation, and deletion are usually additional systems. The five-layer guide to how AI companions work follows those pieces.

Token

A token is a small unit of text processed by a language model. A token may be a whole short word, part of a longer word, punctuation, or another text fragment. Context-window size is measured in tokens, not necessarily messages or words.

Prompt

A prompt is the input supplied to a generative model. In a companion product, the prompt may include much more than the message you typed: system instructions, persona information, recent turns, retrieved memories, safety rules, and technical formatting.

This is why “prompt” and “text message” are not always the same object. You see the text; the model may receive a constructed package.

Hallucination or confabulation

A hallucination is generated information that is false, invented, or unsupported but may sound plausible. NIST prefers confabulation and defines it in a current US-EU AI terminology document.

Companion warmth does not reduce this risk. A familiar voice can make an incorrect date or recommendation feel more credible. Verify anything factual that matters, especially health, legal, financial, safety, or other high-stakes information.

The continuity words

Context window

A context window is the limited amount of information a model can consider while producing a response. It may contain recent messages, instructions, and retrieved information. When a long conversation exceeds what is included, older material may be summarized or left out.

A large context window can improve continuity inside a long session. It is not automatically long-term memory.

Persistent memory

Persistent memory is information stored outside one temporary model response so it can be used in later conversations. A companion might store a name, preference, plan, correction, or summary.

Persistent does not mean permanent, perfect, or exhaustive. A service should explain retention and allow correction or deletion. Read how memory works in an AI friendship before treating “remembers everything” as a meaningful claim.

Memory extraction

Memory extraction is the process of selecting details from a conversation for possible storage. The system may infer that a person, preference, or event will matter later. That inference can be wrong.

For Mia, Warmth’s Privacy Policy says the service extracts details such as names, preferences, and plans, that stored inferences may be inaccurate, and that users can request correction or deletion.

Retrieval

Retrieval is the process of finding stored information that may be relevant to the current message. The research term retrieval-augmented generation describes pairing a generative model with external information; the original RAG paper focused on knowledge tasks, while companion products can use related patterns for personal context.

Storing a detail is only half the problem. The product must retrieve the right detail at the right time and avoid bringing back something irrelevant or deleted.

Proactive messaging

Proactive messaging means the AI can initiate a message rather than waiting for a new prompt. Examples include a morning hello, a follow-up on a plan, or a new conversational topic.

Proactive does not mean unlimited permission. Good systems respect consent, quiet hours, cadence, and silence. The AI friend that texts first guide explains the difference between a considerate check-in and pressure.

The media and trust words

Multimodal

Multimodal means a system can work across more than one type of information, such as text, audio, images, or video. A companion may understand an image, transcribe a voice note, generate a photo, or speak a text reply.

The word alone does not tell you which directions work. “Supports voice” might mean accepting your audio, sending generated audio, offering a live call, or all three.

Synthetic media

Synthetic media is generated or substantially altered text, audio, imagery, or video. For an AI companion, a selfie-like image and a spoken voice can both be synthetic.

The media can support a consistent fictional identity, but it should not be used to imply that a human took a documentary photo or made a recording. Clear disclosure protects the imaginative frame.

Human review

Human review means an authorized person can access some user content for a defined purpose, such as handling a support request, investigating abuse, meeting a legal obligation, or evaluating safety and quality. It is different from a person secretly writing the AI’s replies.

Read the policy for who may review, why, how access is restricted, and whether data is minimized. Warmth describes its current review circumstances in the Privacy Policy.

Model training

Model training can refer to several different practices: building a public foundation model, fine-tuning a provider model, improving a specific product, or evaluating outputs. A policy that says “we do not train on your data” should say which meaning it intends.

Warmth says it does not use conversations to train publicly available foundation models and contractually prohibits model providers from training their own models on them. It may use conversations to operate, evaluate, and improve Mia, including limited safety and quality review with minimization and de-identification where possible.

How to use this glossary

When a product page says “agent,” ask what it can act on. When it says “memory,” ask what is stored and for how long. When it says “voice,” ask whether that means clips or live conversation. When it says “companion,” ask what boundaries distinguish companionship from care. When it presents a photo, ask whether the image is generated.

The vocabulary is useful only when it makes a product testable. Pair this glossary with the AI companion privacy checklist and ten questions for choosing a companion. Clear words will not guarantee a good relationship-style experience, but vague words almost always make one harder to evaluate.