The short version: the model itself remembers nothing. A companion remembers by writing something down and sending it back with every reply — so what it remembers is whatever survived being written down.
A language model holds nothing between calls. Every reply starts from an empty state, and everything the character appears to know arrives as text inside the same request: its own description, whatever the service has stored about you, and some of the recent conversation. Take that text away and it is a stranger again. So the useful question is not whether it remembers, but what this particular service writes down and how much of it gets sent back.
The recent transcript. The last handful of exchanges is re-sent word for word. This is why a companion follows the thread of a conversation you are in the middle of, and why it loses that thread after a gap: the window is short and older turns drop off the end.
A written summary. The service keeps a compressed note about you — a few sentences, or a running list of things you mentioned — and re-sends it with every message. It is kept small deliberately, because everything sent is paid for and slows the reply down. A summary is a lossy copy: whatever did not make it in is gone, however clearly you said it.
Search over stored messages. Some services keep everything and fetch the parts that look related to what you just wrote. That recovers detail a summary would have dropped, but it finds what your wording points at; ask in different words and the same fact can stay buried.
Every remembered word is re-sent on every single reply, and models are billed by the amount of text going in as well as coming out. A companion that carries a long memory therefore costs its operator more per reply than one that carries a short memory. That is the whole reason memory limits exist, and the reason a service under load may quietly send less of your history than usual. If a service tells you how much it keeps, that is a number you can reason about; if it does not, the size of the memory is the part of the product you cannot see.
A memory is a written description of you, held by a company, built without your review. Two questions settle how much that matters: can you read it, and can you erase it. A service that shows you the note is telling you exactly what it holds. A service that erases it on request, immediately and without a backup that quietly restores it, is one where a mistake is fixable. Neither is guaranteed by having a delete button in the interface, because hiding a conversation from the screen and erasing the record are different operations.
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.
What it is, and what it is not
Write to hello@toolkitlabs.org. It is a real inbox on this domain and the messages are read.