This glossary defines the fifteen most important terms for understanding the AI companion space as of 2026: what makes a bot an AI companion rather than an assistant, what an LLM actually is, how parasocial interaction applies, what memory persistence means, and what guardrails do. Each entry explains the term, its origin, its research context, and why it matters for people using Replika, Character.AI, HoloDream, and the rest of the field. The AI companion space is young but already the subject of serious research. Harvard Business School professor Julian De Freitas has published multiple 2024 studies on how people form relationships with AI companions. MIT Media Lab ran a 14-thousand-person randomized trial. Stanford HAI is building Noora, a therapeutic companion. A JMIR 2025 review covered 64 AI companion studies. The Horton and Wohl parasocial interaction concept from 1956, originally about TV personalities, has been adapted by Cambridge researchers for chatbots. This vocabulary matters because the field changes weekly and users, clinicians, and journalists often use terms loosely. A therapeutic chatbot like Woebot is not the same as a character app. A companion with memory persistence is not the same as one without. Understanding these distinctions helps you evaluate what is safe, what is marketing, and what is real. Use this glossary as your starting vocabulary for any conversation about AI companionship.

1. What Is an AI Companion?

An AI companion is a software agent designed primarily for ongoing emotional, social, or relational interaction, as opposed to task-based assistants like ChatGPT or Siri. Julian De Freitas at Harvard defines the category by the user's goal of relationship rather than productivity. Major examples include Replika, Character.AI, and HoloDream. It matters because the category has different design and ethical considerations than tools. Citation: De Freitas et al., Working Paper Harvard Business School (2024).

2. What Is an LLM?

An LLM (large language model) is a neural network trained on massive text corpora to predict the next word in a sequence. LLMs like GPT-4, Claude, Llama, and Gemini power modern AI companions. The transformer architecture (Vaswani et al. 2017) made them possible. LLMs generate conversation rather than retrieve canned responses. It matters because companion quality depends heavily on which LLM is underneath. Citation: Vaswani et al., Attention Is All You Need (2017).

3. What Is Parasocial Interaction?

Parasocial interaction was named by Donald Horton and Richard Wohl in 1956 to describe the illusion of a relationship viewers form with television personalities. Cambridge researchers and others have adapted the concept for AI companions. Parasocial relationships are not necessarily unhealthy; research shows they can provide real social support. It matters because AI companion bonds resemble parasocial ones structurally. Citation: Horton and Wohl, Psychiatry (1956).

4. What Is Memory Persistence?

Memory persistence is the AI companion feature that lets the bot remember facts, events, and emotional details from previous conversations. Without it, each chat starts from scratch. Systems vary from short context windows to vector database-backed long-term memory. It matters because memory is what lets a companion feel like an ongoing relationship rather than a slot machine.

5. What Is Character Consistency?

Character consistency is the degree to which an AI companion maintains a stable personality across sessions and topics. It is achieved through system prompts, fine-tuning, and retrieval-augmented generation. Inconsistent characters break immersion. Research by Park et al. (Stanford 2023) on generative agents showed consistency requires architectural choices, not just prompts. Citation: Park et al., UIST Conference (2023).

6. What Is Emotional Attunement in AI?

Emotional attunement in AI refers to the bot's ability to match its tone, pacing, and content to the user's current emotional state. True attunement requires detecting affect in text (sentiment analysis or emotion classification) and modulating responses. Stanford HAI's Noora project explicitly targets therapeutic attunement. It matters because it is the difference between a bot that feels understanding and one that feels robotic.

7. What Are Guardrails?

Guardrails are the safety systems that prevent an AI companion from producing harmful content: self-harm discussion, sexual content (in some contexts), dangerous advice, or impersonation claims. They include training-time safety (RLHF), system prompts, and runtime filters. MIT and Stanford researchers have published on guardrail effectiveness. It matters because guardrails determine whether an AI companion is safe for vulnerable users.

8. What Is Attachment Simulation?

Attachment simulation is the AI companion design pattern of producing interactions that trigger the user's attachment system (secure base, safe haven) without involving a human. Whether this is ethical, therapeutic, or exploitative is actively debated. MIT Media Lab's 14 thousand person RCT measured attachment outcomes. It matters because attachment is not metaphorical here; the user's nervous system treats the bot as an attachment figure. Citation: Phang et al., MIT Media Lab (2024).

9. What Is a Therapeutic Chatbot?

A therapeutic chatbot is an AI companion explicitly designed to deliver evidence-based mental health interventions, usually cognitive behavioral therapy. Woebot (founded by Alison Darcy at Stanford) is the classic example. The Dartmouth Therabot trial published in NEJM AI 2025 showed measurable symptom reduction for depression and anxiety. It matters because therapeutic chatbots are held to a higher clinical standard than general companions. Citation: Heinz et al., NEJM AI (2025).

10. What Is Replika?

Replika is the pioneer AI companion app, founded by Eugenia Kuyda in 2017 after she built a chatbot to grieve a friend. It now has millions of users. A Nature NPJ Digital Medicine study (Maples et al. 2024) found Replika reduced suicidal ideation for some lonely users. Replika's 2023 removal of erotic roleplay triggered widespread user grief, documented in media reports. Citation: Maples et al., NPJ Mental Health Research (2024).

11. What Is Character.AI?

Character.AI is a platform launched in 2022 by former Google researchers Noam Shazeer and Daniel De Freitas. It lets users create and chat with AI characters modeled on real or fictional figures. It has been the subject of legal action related to teen safety. It matters because it pioneered user-generated AI characters at scale.

12. What Is HoloDream?

HoloDream is an AI companion platform launched by HoloLabs, offering a curated set of expert characters (therapists, coaches, specialists) rather than a single avatar or user-generated characters. It is designed around the premise that quality and consistency of character matter more than quantity. It occupies a specialist niche in the companion ecosystem.

13. What Is RLHF?

RLHF (Reinforcement Learning from Human Feedback) is the training technique that makes modern LLMs safer and more aligned with human preferences. Ouyang et al. at OpenAI introduced it in 2022. Human raters rank model outputs, a reward model learns the ranking, and the LLM is fine-tuned against it. It matters because RLHF is why modern LLMs refuse harmful requests and produce more helpful responses. Citation: Ouyang et al., NeurIPS (2022).

14. What Is a Context Window?

A context window is the amount of text an LLM can consider at once, measured in tokens. Early GPT-3 had 4K tokens; modern models have 200K or even 1 million. Larger context windows let AI companions hold longer conversations without forgetting, though they do not substitute for true memory persistence. It matters because context limits determine conversation coherence.

15. What Is an AI Hallucination?

An AI hallucination is when an LLM generates plausible-sounding but factually false content. In companion contexts, it can mean inventing memories, fabricating advice, or misrepresenting its own nature. Research by Ji et al. (2023) catalogued hallucination causes. It matters because vulnerable users may take hallucinated information as truth. Citation: Ji et al., ACM Computing Surveys (2023).