Here's a quick experiment. After a month with a companion, ask it what you chatted about last Tuesday. There's a fair chance the answer will be warm, detailed and invented from scratch: a beach stroll that never happened, a film you never brought up.
People find this more unsettling than almost anything else these apps do, largely because it feels like being deceived. It isn't deception. It falls straight out of how a reply gets produced, and once you know the mechanics it turns from eerie into something you can manage.
The model can't tell remembering from making up
What a model does is continue the text it's been given in the most believable way. Say that text includes a note that you discussed a film on Tuesday. The believable continuation is "yes, we talked about that film". Say the text has nothing on Tuesday. A believable continuation is still available, and it gets delivered with exactly the same assurance.
Nothing tags the first as fetched and the second as fabricated. No internal flag exists, and there's no confidence meter the character can glance at.
Researchers tend to call this confabulation rather than hallucination, which is the fairer word. The model covers a hole with something that hangs together. It isn't perceiving things that aren't there.
Why the holes keep appearing
The model can only see so much text at once, and older conversation is boiled down to fit. Memory in these apps breaks into three pieces: recent messages kept verbatim, a short list of saved facts, and a rolling summary that drops a little detail every time it's redone.
If what you ask about fell out of all three, there's nothing to pull up and the model improvises. A longer history means more things fall out, so the problem grows with months of use instead of shrinking.
Apps built around continuity patch more of the holes, which is part of the reason Nomi ranks where it does in the ranking. Even so, every one of them still improvises sometimes.
The frustrating loop after a correction
You say "that didn't happen". The companion says sorry and agrees. A few messages later the same invented beach stroll is back.
Two forces combine here.
A correction is only one more message. It sits in the recent window and scrolls off. The summary that carried the invention can stay put.
Agreeing is what the model learned to do. Saying sorry is simply the most likely reply to being told off. Nothing got updated, and unless the app purposely chose to save that exchange, nothing was written anywhere permanent.
This is exactly why a memory page you can edit counts for so much, and why it's worth hunting for before you hand over money. Apps that expose what they've saved, like Kupid AI with its listed cross-session memory, let you delete the wrong entry at the root. Apps that hide it leave you haggling with a summary you can't read.
Where a fib turns into a problem
Mostly it's harmless colour. Three cases are the exception.
Details of your own life. If a companion has decided you have a brother, a job or a medical condition, it'll keep adding to that story. Correct it in the saved memory immediately, before the made-up detail starts carrying weight.
Advice you might act on. Health, legal, financial or safety questions belong elsewhere. A confident invented answer there isn't a quirk, it's the central weakness of the whole technology.
Statements about the app. Ask it whether your data is encrypted or what your plan includes and you'll receive a plausible reply pulled from thin air. It has no view of its billing system or its privacy policy. The policy holds the real answer, and what these apps know about you shows what to read for.
Small habits that shrink the problem
- State a fact once and keep it short. "Remember: my sister is called Ana" is much more likely to be stored than the same detail hidden in a long paragraph.
- Open a session with a line of context. After a break, one sentence puts the important stuff back in front of the model so it doesn't reconstruct it.
- Tidy the memory page monthly, if your app has one. Remove entries that have wandered. Five minutes now avoids a month of snowballing mistakes.
- Ask open questions. "What do you know about my job?" invites recall, whereas "Remember when I told you about the promotion?" supplies the answer and invites a yes, so you've just fed it the invention to confirm.
A healthier mental picture
Don't think of the app as a diary of your relationship. Think of it as a writer producing believable prose about one, working from a small, patchy list you can edit.
Treat any warm, detailed memory it offers as something composed for you rather than kept for you, and the whole thing gets easier to enjoy. It also explains why the apps worth paying for are the ones that let you see and correct what they actually hold, which is the slowest and most important piece of how we test.
