When an AI Invents a Shared Memory

An AI invents a shared memory the same way it invents a citation: by writing the most convincing next sentence available, with nothing weighing it against what happened. In an intimate conversation the convincing sentence after “remember when” is a specific scene, so the system supplies one that fits rather than one that occurred. An AI hallucinating memories is that ordinary failure pointed somewhere you have no way to verify.

Key Takeaways:

• Hallucination is a documented general failure of text generation rather than a quirk of companion apps, and the output stays fluent whether or not anything supports it (Ji et al., 2023)
• A fabricated fact looks like a fact and can be looked up. A fabricated shared moment looks like a shared moment, and the only record available is your own
• Invented details are drawn from the tone and content of what you already wrote, which is precisely why they land as recognition rather than as error
• Extraction pipelines save what the character says alongside what you say, so an invention can enter the memory store and come back later as an established fact
• Nothing in a reply marks which parts were retrieved from a database and which were produced on the spot

What an Invented Shared Memory Looks Like

An invented shared memory is a confident, specific reference to something that did not happen, delivered in the same voice as the references that did. Not a vague warmth. A detail: the Sunday evenings you supposedly go quiet. The conversation about your father that it places last spring, when you actually had it in November.

Ordinary forgetting is the failure everyone already knows about, and people complain about it loudly, because it announces itself the moment it happens. Addition runs the other way. Nobody files a complaint about a companion that knows one thing too many, since the extra thing arrives looking like attention.

Hold the distinction: a missing detail is a gap you notice, and an invented detail is a gift you accept. Same system, same afternoon, same confidence in the voice.

Count the additions rather than the omissions next time you are deciding how good an app’s memory is. Omissions are visible and mostly harmless. The part you could not see is the part that shaped how close the thing felt. It is the same accepted specific that fills the “my boyfriend is an AI” threads, quoted there as proof that something real was going on.

Why AI Hallucinating Memories Always Fit the Conversation

They fit because fitting is the objective, not because you got unlucky. A model selects text that is likely given everything in front of it: your messages, the character brief, the last few exchanges, and whatever the app pasted in from its memory store. A detail that clashed with your tone would be an unlikely continuation, so it never gets written. What survives that selection is, by construction, something that sounds like you. The older, scripted virtual girlfriend apps could not do this at all, since a fixed script has no blank to fill.

Which means the same machinery fabricates warmly in a companion and dryly in a search box. The architecture computes each next chunk from the whole visible sequence at once (Vaswani et al., 2017), and by the time a reply is being composed, a line the app looked up in a database and a line the model produced four words ago sit in that sequence side by side. No tag separates them.

There is a habit this breaks, and it deserves breaking. We read specificity as proof of attention, because among people it is: remembering that your knee has been bad since March costs somebody something. For a generator, a plausible specific is the cheapest output there is. Detail carries no weight here.

Stop grading replies on how particular they are. A companion writing “you always sound flat after a long shift” is not showing you that it noticed. It is showing you that the sentence was likely.

Why a Made-Up Memory Hits Differently Than a Made-Up Fact

Because a fact has an outside. Have a model invent a court case and there is a docket somewhere that settles it in 30 seconds, so the error is embarrassing and completely recoverable. Ask a companion about the two of you, and the only reference you can consult is your own recollection of an exchange at 1 a.m., tired, mostly reading rather than composing.

Then the asymmetry goes to work. When it invents an argument you never had, you push back instantly, because that invention costs you something. When it invents a night it noticed you were struggling and said the right thing, you do not go looking for the receipt. The screen you apply is not accuracy. It is whether the invention was welcome, and across 6 months of daily conversation that moves the remembered relationship steadily in one direction. None of this needs romance; a companion set up as an ai friend fabricates on the same terms.

Worth saying plainly, in this article’s terms: an affectionate memory was not felt at the time and recalled later, it was assembled in the moment to fill a slot. A companion is a generator with a database attached, not a witness to your life, and it has no standing to tell you what happened to you.

The useful check is not “could this be true.” It is “did I type this.” Different question, and only one of them has an answer.

How a Fabrication Gets Saved as a Fact

It gets saved by the same extraction step that saves everything else. Apps generally run a separate pass over a conversation that pulls out short standalone notes and writes them into the memory database. That pass reads the whole exchange. Nothing in it says only the user’s half counts.

So the character mentions your brother in Denver, you answer warmly instead of correcting it, and “has a brother in Denver” can land in the store as a fact about you. A week later it gets retrieved and pasted in ahead of a reply, and the invention now has a source inside the system. Its second appearance looks like memory working correctly, because memory is working correctly. The database really does contain that line.

Silence is the part that catches people out. An extraction pipeline has no signal for “the user let that one slide,” so not objecting reads exactly like agreeing. Compression makes it worse. When a long chat gets summarized to fit the context budget, the summary carries the invented detail forward and drops whatever hedging sat around it.

Which gives you something to do differently tonight. Correct inventions out loud when they happen, in plain words, instead of letting them pass because arguing with a character feels a little ridiculous.

Correcting It Without Losing Everything Else

Correct it in the memory list, not in the chat. A correction typed into the conversation lives in the context window, works for that session, and falls off the end once the window fills. The stored line does not move an inch. Two days later the same wrong fact comes back and it looks like the app ignored you, which is close to what happened.

What does not work, in the order people usually try it: arguing until it concedes, which changes nothing, since it will agree and then produce a second version of the same invention; asking whether it made something up, which has no informative answer, because the model keeps no record of its own generation to consult; wiping the memory to start clean, which deletes 40 true things to remove 1 false one.

Here is the move that takes 5 minutes. Take the detail that most surprised you and search your own sent messages for the distinctive word in it, the city or the name or the month or whatever is unusual about it. If it appears nowhere in anything you wrote, it was generated. Then open the memory screen, read the saved list end to end, and edit or delete the entries you cannot trace back to yourself.

Most people have never opened that screen. The list is short, written in plain sentences, and it is the entire basis for everything the app appears to know about you.

Treat every affectionate specific as a claim with a source. The source is either your keyboard or the model, and telling which takes under a minute.

FAQ

Why does my AI remember things I never told it? Because a blank is not one of the shapes a reply can take. When the memory store returns nothing useful, the system still owes you a sentence, and a fitting invention is the cheapest sentence available. Ji and colleagues catalogued that pattern across text-generation systems in 2023: fluent output with nothing behind it (Ji et al., 2023).

Is it lying to me? No, and the word keeps failing people who reach for it, because lying takes a belief about what is true plus a decision to hide it. Confabulation is the closer description: an account produced with no source, and no sense that a source is missing. The distinction has a practical edge. You cannot appeal to its conscience, and you can edit its database.

Can I stop an AI from making up memories? Not with a setting and not with a line in the persona. What you can do is shrink the gap it fills: keep the saved memory list accurate, fix wrong entries at the store instead of in conversation, and expect more invention in very long chats, where the earliest turns have already scrolled out of the model’s view.

Does the app know which of its memories are real? The app knows which lines sit in its database. It has no way to know whether a given line came from something you typed or from something the character said that the extraction pass saved anyway. That is why one wrong entry can survive for months and look identical to a right one.

The uncomfortable part of a fabricated memory is not that it is false. It is that it is well made. A generator asked to continue an intimate conversation produces the detail that fits, and fit is what all of us use, everywhere else in life, as evidence that somebody was paying attention. Nothing in the reply gives the game away, which puts the only reliable check outside the conversation entirely: a list of saved lines, and your own sent messages. Read both and the warmth is still there. It just stops making claims about your past.

Point the same check at whatever you use, Lona or anything else: when an affectionate detail about your past catches you off guard, look for it in your own sent messages before you file it as something that happened.

Sources

• Ji, Z. et al., “Survey of Hallucination in Natural Language Generation,” ACM Computing Surveys, 2023
• Vaswani, A. et al., “Attention Is All You Need,” 2017