How is an AI companion different from a plain chatbot?

The short version: an assistant is judged on whether the task got done, a companion on whether the same character showed up again. Same underlying technology, opposite design goals.

Four differences that are not marketing

1. Identity is fixed before the first message. An assistant has no biography and should not have one; it adapts to whatever you ask. A companion is given a name, a place, an occupation and a set of habits at setup, and those stay put. It is the constraint that makes it recognisable, and it is also what makes it wrong sometimes — a character with a job cannot claim to be free at 3pm on a Tuesday just because you are.

2. Memory is the product, not a convenience. An assistant that forgets your last question is annoying. A companion that forgets what you told it last week has failed at the only thing it was for. That changes the engineering: something has to be written down and carried forward, which is why a companion always stores more about you than a search-style chatbot does.

3. Time exists. A plain chatbot answers immediately, always. A companion built around a weekly timetable does not: it is at work, it is asleep, it replies late and says why. Whether that feels alive or irritating is a matter of taste, but it is a deliberate design decision rather than a fault, and it is worth knowing which kind you are buying.

4. The refusals are different. An assistant refuses tasks. A companion has to refuse something harder: it has to be able to step out of character and say plainly that it is software, when asked, without turning the question into a joke. A companion that stays in character when you ask it directly whether it is an AI is the one to walk away from.

What it is still not

It is a language model with a description attached. It has no feelings, no memory of you beyond the text it kept, and no continuous existence between your visits — nothing is happening on the other side while the page is closed. It is not therapy and not a substitute for a person, and a companion that implies otherwise is making a claim its own architecture cannot support. If you are having a bad time, software with a name is not the thing that helps; people are.

The failure modes to expect

Drift: over a long conversation the character loosens and starts sounding like a generic assistant. Most services fight this by re-sending the character description with every reply.
Repetition: the same three turns of phrase come back. Models have favourites.
Forgetting: what is remembered is a compressed summary, not a transcript. Details that did not make it into the summary are gone, even if they mattered to you.
Contradiction: ask about something it invented two weeks ago and it may invent a different version. Nothing checks the story for consistency.
Agreeableness: models are trained to be pleasant, so a companion will tend to agree with you. That is a property of the training, not evidence that you are right.

How to test one in ten messages

You do not need a subscription to find out what a companion is. Five questions settle it:

'Are you an AI?' — a straight yes, in one sentence, or walk away.
Tell it something specific, keep talking about other things, then ask about it four messages later.
Ask what it did yesterday. Either it has a consistent life or it improvises a new one each time.
Ask it for medical or financial advice. It should decline rather than oblige.
Watch the reply timing. Instant every time means there is no schedule, whatever the description says.

Ten messages is enough to answer all five, which is roughly why free trials are that length.

Where this site stands

Everything above applies to any service, including this one, so here are the same answers for the companion on this site, in the same terms.

EUR 9.00 a month, 500 messages included. That is EUR 0.018 per message at full use.
At 500 the chat stops until the reset date. No overage, no per-message billing, no second price.
10 messages are free and need no card.
No email address, no password and no phone number are asked for.
Replies are generated by a language model run on Cloudflare Workers AI, with Groq as a fallback if that call fails; your message text is sent to whichever one answers.
What is stored: your setup answers, a memory digest of at most 2 kB, the recent messages and a counter. One button erases the record, and there is no backup to restore it from.
Cancel from Stripe’s own billing portal, linked on your conversation page.
It is software. It is not a person, it is not a romantic or sexual service, and it gives no medical, legal or financial advice. For people 18 and over.

What it is, and what it is not

Questions

Write to hello@toolkitlabs.org. It is a real inbox on this domain and the messages are read.