Try this after a month with a companion: ask what the two of you discussed last Tuesday. Odds are good you'll get a warm, detailed and completely invented reply, like a walk you never took or a movie you never brought up.
Few things in this category rattle people more, mostly because it feels like being lied to. It isn't. It follows directly from how replies get built, and once you understand that, it's less spooky and easier to manage.
Recalling and inventing are the same act
A model writes the most believable next stretch of text given what's in front of it. Suppose a note says you talked about a film on Tuesday. Then "we talked about that film" is the believable next line. Suppose there's no note about Tuesday at all. A believable next line still exists, and it comes out with identical confidence.
Nothing in the process labels one answer as looked up and the other as made up. There's no internal flag and no confidence gauge for the character to check. That's the entire explanation.
Researchers often prefer the word confabulation to "hallucination," and it fits better. The model is patching a hole with something coherent. It isn't seeing something that isn't there.
Holes are everywhere
The context window has a size limit, and older history gets squeezed down. Memory in these apps is really three separate layers: recent messages kept word for word, a short list of saved facts, and a running summary that loses detail every time it's rewritten.
Ask about something that dropped out of all three and there's nothing to fetch, so the hole gets filled. More history means more holes. That's why the problem gets worse the longer you've used an app, not better.
Apps designed around continuity handle holes better. That's part of why Nomi scores the way it does in our ranking. Still, none of them get rid of holes entirely.
Why saying "that didn't happen" rarely sticks
You push back. It says sorry, agrees, and three messages on it mentions the same invented walk.
Two reasons.
Your correction lives only in the window. It's a recent message, so it will scroll away. The summary that carried the invention might not.
Agreeing is what it was trained to do. An apology is simply the likeliest reply to being contradicted. Nothing got updated. Unless the app specifically chose to save that exchange, nothing was written to storage.
Hence the value of an editable memory screen, and why you should check for one before subscribing. Apps that show their stored facts, such as Kupid AI, which lists cross-session memory among its features, let you delete a bad entry at its source. Apps that hide it leave you arguing with a summary you can't see.
When it stops being cute
Most confabulation is harmless flavor. Three kinds aren't:
Facts about your life. Once a companion invents a sibling, a job or a diagnosis, it keeps building on it. Fix it at the source right away, or the invention starts holding weight.
Anything you might act on. Medical, legal, money or safety questions. A confident fabricated answer here isn't a quirk. It's how this whole technology fails, and no companion app is the right tool for those questions.
Statements about the app itself. Ask whether your data is encrypted or what your plan includes, and you'll get a plausible answer built from nothing. The character can't see its own billing system or privacy policy. The real answer is in the policy, and what these apps know about you explains what to look for.
Four habits that help
Say facts once, plainly. "Remember: my sister's name is Ana" is much likelier to be saved than the same fact tucked inside a paragraph.
Reset the scene after time away. One sentence of context at the top of a session puts what matters back in front of the model, so it doesn't have to guess.
Review the memory screen monthly, if there is one. Delete entries that have wandered off. Five minutes now heads off a month of compounding errors.
Ask open questions. "What do you remember about my job?" invites recall. "Remember when I told you about my promotion?" hands over the answer and invites a yes, so you've just given it the invention to confirm.
A better way to think about it
A companion app isn't a record of your relationship. It's a system that writes believable text about one, resting on a small, lossy, editable list.
Read the warm, detailed memory it offers as something it wrote for you, not something it kept for you, and the whole experience gets easier to enjoy. It also shows why the apps worth paying for let you see and fix what they actually store. That's the slowest and most important part of our testing.
