Pick a companion you have chatted with for a month and ask what you talked about last Tuesday. The reply may be affectionate, specific and entirely fictional: a stroll you never took, a movie nobody mentioned.
For many users this unsettles more than anything else about these apps, mainly because it looks like dishonesty. It is nothing of the kind. It flows straight from the way replies are composed, and understanding that takes away the eeriness and makes it manageable.
Remembering and inventing look identical to the model
A language model extends whatever text it receives in the most probable way. Given a note saying you spoke about a movie on Tuesday, "we discussed that movie" is the obvious next step. Given nothing about Tuesday, a plausible next step still exists, and it comes out just as confidently.
Nothing in the process marks one answer as looked up and the other as made up. The character has no internal tag to read and no certainty meter to check. That covers the whole explanation.
Many researchers favour the term confabulation over hallucination, and rightly. The model fills a gap with something coherent; it does not perceive anything that is not there.
Why there are always holes
Only so much text fits in the model's view, so older history is squeezed down. Memory in these apps has three layers: recent messages held verbatim, a short list of stored facts, and a rolling summary that drops detail with every rewrite.
Ask about anything that fell out of all three layers and there is nothing to retrieve, so the model fills in. Longer history brings more gaps, which is why things tend to worsen, not improve, on an app you have used for months.
Apps designed for continuity, which is why Nomi scores as it does in our ranking, handle the gaps better. None removes them.
Why a correction often does not stick
You say it never happened. The companion apologizes and agrees, then three messages later mentions the same imaginary walk.
Two things are at work.
The correction sits only in the current window. It is one recent message and will scroll out of view, while the summary that captured the invention may stay put.
Agreeing is what the model is trained to do. Apologizing is simply the likely reply to being contradicted; it does not mean anything was updated. The exchange writes nothing to storage unless the app has been built to save it.
This is why an editable memory screen matters so much, and why it is worth checking before you subscribe. When an app shows its stored facts, as Kupid AI does with cross-session memory among its features, you can delete the wrong entry at its origin. When it hides them, you end up arguing with a summary you cannot read.
When it stops being harmless
Most confabulation is innocent colour. Three kinds are not:
Details of your own life. A companion that has made up a sibling, an employer or a diagnosis will keep building on it. Fix these in the stored memory straight away, or the invention becomes something the rest rests on.
Anything you could act on. Take medical, legal, financial and safety questions elsewhere. A self-assured invented answer there is the basic weakness of the technology and not a charming quirk, and a companion app is the wrong place to ask.
Claims about the app itself. Ask whether your data is encrypted or what your plan covers and you will get a believable answer built from nothing, since the character has no view of billing or of the privacy policy. The truth is in the policy; what these apps know about you shows what to read for.
Four habits that help
State facts once, simply. A line such as "Remember: my sister is called Ana" is far likelier to be stored than the same detail buried mid-paragraph.
Re-establish context after a gap. A sentence at the start of a session puts what matters back where the model can see it and spares it from reconstructing the details.
Review the memory screen every month, where one exists. Remove entries that have drifted. Five minutes now prevents a month of errors piling up.
Prefer open questions. "What do you recall about my job?" prompts retrieval. "Remember when I told you about the promotion?" supplies the answer and requests a yes, giving it something invented to confirm.
A healthier way to see it
A companion app is no diary of your relationship. It generates convincing prose about one, supported by a small, incomplete list you are able to edit.
Treat any warm, detailed recollection as something it wrote for you rather than kept for you, and the whole experience sits more comfortably. It also explains why the apps to pay for are those that let you inspect and correct what they genuinely store. That is the part of our testing that takes longest and counts most.
