Ask a companion you have used for a month what you talked about last Tuesday. There is a good chance you will get a warm, specific, entirely fabricated answer — the walk you did not take, the film you did not mention.
This unsettles people more than almost anything else in the category, partly because it looks like dishonesty. It is not. It is the direct consequence of how the reply is produced, and once you see the mechanism it stops being eerie and becomes manageable.
There is no boundary between recalling and inventing
A model produces the most plausible continuation of the text in front of it. If the text contains a note saying you discussed a film on Tuesday, "we talked about that film" is the plausible continuation. If the text contains nothing about Tuesday, a plausible continuation still exists — and it will be generated with exactly the same confidence.
Nothing in the process tags one as retrieved and the other as manufactured. There is no internal flag, no confidence meter the character consults. That is the whole of it.
The technical term people prefer is confabulation rather than hallucination, and it is the better word: the model is filling a gap with something coherent, not perceiving something that is not there.
Why gaps appear constantly
They appear because the context window is finite and history is compressed. Memory in these apps is three systems — recent messages held verbatim, a short list of stored facts, and a rolling summary that loses detail every time it is rewritten.
Ask about something that fell out of all three and there is nothing to retrieve. The gap gets filled. The more history you have, the more gaps exist, which is why this gets worse on an app you have used for months rather than better.
Apps built for continuity — the reason Nomi scores as it does in our ranking — manage the gaps better. None of them eliminate them.
Why correcting it often fails
You say: that never happened. It apologises, agrees — and three messages later refers to the same invented walk.
Two things are going on.
The correction is only in the window. Your denial is a recent message. It will scroll out. The summary that contained the invention may not.
Agreement is the trained behaviour. The model apologises because apologising is the plausible continuation of being contradicted, not because anything was updated. Nothing was written to storage by that exchange unless the app specifically decided to store it.
This is the practical reason an editable memory screen matters so much, and why it is worth checking before you subscribe. Apps that expose stored facts — Kupid AI lists cross-session memory among its features — let you delete the wrong entry at the source. Apps that hide it leave you arguing with a summary you cannot see.
Where it stops being charming
Most confabulation is harmless colour. Three cases are not:
Facts about your life. A companion that has invented a sibling, a job or a diagnosis will keep building on it. Correct these at the source immediately, or the invention becomes load-bearing.
Anything actionable. Medical, legal, financial or safety questions. A confident invented answer here is not a quirk; it is the failure mode of the whole technology, and no companion app is the right tool.
Claims about the app itself. Ask a companion whether your data is encrypted or what your subscription includes and you will get a plausible answer generated from nothing. It has no access to its own billing system or privacy policy. The answer is in the policy, and we cover what to look for in what these apps know about you.
Four habits that reduce it
State facts plainly and once. "Remember: my sister is called Ana" is far more likely to be stored than the same fact buried in a paragraph.
Re-anchor after a break. A sentence of context at the start of a session puts the things you care about back where the model can see them, instead of leaving it to reconstruct them.
Audit the memory screen monthly if the app has one. Delete the entries that have drifted. Five minutes prevents a month of compounding.
Ask open, not leading. "What do you remember about my job?" invites recall. "Remember when I told you about the promotion?" supplies the answer and invites agreement — you have just handed it the invention to confirm.
The reframe that helps
A companion app is not a record of your relationship. It is a system that produces plausible text about one, and the record is a small, lossy, editable list underneath.
Treat the warm specific memory it offers as something it wrote for you rather than something it kept for you, and the whole thing becomes easier to enjoy. It also makes it obvious why the apps worth paying for are the ones that let you see and correct what they actually store — which is the part of our testing that takes the longest and matters most.
