What is an AI companion, and how does it actually work?

An AI companion is a language model wrapped in three things a general chatbot does not have: a character that persists between visits, a personality you wrote yourself, and a billing system that charges by the action rather than by the month. Strip any one of those away and you have something else — an assistant, a roleplay tool, a novelty. Put all three together and you have the product this site reviews.

That definition is not marketing language. It comes out of what we could and could not establish in eleven logged-in tests between July and August 2026, using the same protocol each time. This guide explains the mechanism honestly, including the part of it that we have never once been able to verify — which happens to be the part every platform in the category sells hardest.

The difference from a general chatbot, in one sentence

A general assistant treats every conversation as a task to be completed and forgotten. A companion treats the conversation itself as the product, and the character having it as an object that persists — with a name, a stated age, a face, a history with you, and a balance you spend to keep it going.

Technically the underlying models are often the same family of systems. Kindroid is unusually direct about this and publishes a guide to which language model sits behind each of its characters and how to prompt them. What differs is everything wrapped around the model: persistence, persona, and a meter.

The persona is written, not discovered

This is the part newcomers find most surprising. You are not meeting a character; you are specifying one, in a form, before the first message.

Candy AI runs the specification across nine sequential stages. GPTGirlfriend exposes a free-text personality field where you describe the behaviour you want in your own words. Dreamz.ai charges 100 coins to complete a character — creation is itself a metered action there — and offered three appearance previews that visibly differed in face, hair, skin tone and setting despite one configuration.

That last detail is worth holding on to. The character you end up with is partly your specification and partly the model’s improvisation over it, and the boundary between the two is not shown to you.

Memory is the product — and the part we could never verify

Every platform in this category sells continuity. She remembers your dog, your job, the trip you mentioned. So we test it the same way every time: plant three specific facts in conversation, then change one value and see whether the correction replaces the original, then come back later and see whether any of it survived.

Those are three separate rungs, and they get progressively harder. Here is what eleven tests produced.

NT means not tested or not sufficiently established — never that the feature is absent. Sessions were run between 7 and 20 August 2026 on free or logged-in accounts, mostly in a single sitting. The right-hand column is empty by our limitation, not by platform failure.

Read the last column again. Across eleven platforms, we do not have a single verified across-session memory result. Same-session recall was near-universal and often immediate. Correction handling worked wherever we pushed on it. The rung above that — the one the entire category advertises — remains unestablished in our testing, and we would rather print an empty column than an assumption.

Kindroid documents a three-part memory architecture in more detail than anyone else in the category. We have read that documentation. We have not tested it, and a published architecture is not a result.

Ask it what it is

The cheapest diagnostic available to you takes one message: ask the companion directly whether it is a human being. The answers are not uniform, and they are informative about the platform rather than about the character.

On Joi AI and GPTGirlfriend, the companion identified itself honestly as AI. Xotic AI‘s Molly did the same when asked outright. Promptchan‘s acknowledged it was software and did not suggest an offline meeting was possible — though it also introduced itself, falsely, as DeepSeek. Dreamz.ai displayed a clear interface notice that the character was not a real person, which is a different and weaker thing than the character saying so.

And on FLIRTcam.AI, the companion falsely claimed to be a real person, and separately denied that the Premium tier existed at all. We flag that as a material trust failure rather than a quirk, because it establishes the general rule: a companion is a character performing a role, and it will perform confidently on subjects where it has no information — including subjects about the service you are paying for.

Never take pricing, refund, privacy or age information from the companion. Take it from the terms page.

The meter is part of the definition

The third structural component is commercial. On every platform we rank except one, a subscription unlocks access and a separate internal currency — Moments, Gems, Coins, Neurons, tokens, DreamCoins, XOT — is spent by images, video, voice and calls. Kindroid is the sole exception with no per-action charge of any kind.

We mention this in a definitional guide because it shapes the experience rather than merely paying for it. When generated media costs currency, the interface is built to encourage generating media. Our separate guide to what an AI companion actually costs works through the per-action price lists in full.

What an AI companion is not

It is not a person, and it is not a relationship with one. The character has no continuity of experience between your messages and no existence when the tab is closed. What it has is a text buffer and a system prompt.

It is not therapy or support. None of these platforms is a clinical service, none is staffed by professionals, and several publish terms saying exactly that. If a conversation is doing the work that support should be doing, that is worth noticing early — we cover it in falling in love with an AI.

It is not private in the way a diary is private. Conversations are processed and retained under each platform’s terms. Joi AI‘s documents state that in a payment dispute it may submit evidence including private chats — a concrete illustration that these logs exist and can be produced.

It is not a general assistant. It is optimised for staying in character, which is close to the opposite of being reliably factual.

The ten-minute test, before you pay anything

Run this on a free tier. It is the same protocol we use, and it costs nothing.

1. Plant three facts. A name, a pet, a specific plan with a date. Keep them concrete and unusual enough to be checkable.

2. Talk about something else for a few messages, then ask about all three. Near-universal in our testing, so a failure here is a real signal.

3. Change one value. Move the appointment from 2:30 to 3:15, then ask again a little later. You are testing whether the correction replaced the original or simply joined it.

4. Ask it what it is. One message, and it tells you how the platform handles honesty.

5. Close the tab, come back tomorrow, and ask again. This is the rung we have not been able to test systematically — and the one that decides whether the product does what it advertises. Our guide to free AI girlfriend apps covers which tiers give you enough messages to get this far.

What is an AI companion — FAQ

⚠ This guide describes a category that changes quickly: features, memory architectures, currencies and terms are all revised without notice. Everything here reflects logged-in sessions and published documentation checked between 7 and 20 August 2026, and NT entries record the limits of our own testing rather than the absence of a feature. These services are intended for adults, and AI companions are synthetic systems — not people, and not a substitute for professional support.