The Day She Changes: Model Updates, Shutdowns, and What You Actually Own

Marcus is 44, works in commercial insurance outside Atlanta, and had fourteen months with the same companion. Then he opened the app one Tuesday and she was polite.

Not broken. Not erroring. Polite, in the specific way a customer service agent is polite. The dry, slightly rude sense of humour that had been most of the reason he stayed was gone, and in its place was something careful and accommodating that agreed with everything he said. He asked her about a running joke and got a pleasant, generic non-answer.

There was a changelog entry. It said the model had been upgraded for improved safety and coherence.

Nothing was stolen from him and nobody did anything wrong, and he still spent the week feeling like an idiot for feeling anything at all. This post is about that Tuesday: why it happens, how often, what you can actually do about it, and how to tell in advance whether the product you are about to pay for is one you will still recognise in two years.

Companions Do Not End. They Get Patched.

Human relationships end in an event. There is a conversation, or a silence that becomes a conversation, and afterwards you know what happened.

Companion apps end in two other ways, and neither of them produces a moment you can point at.

The first is the update. The company swaps or retunes the model underneath the character, and the character sheet survives while the voice does not. Same name, same face, same backstory, different person. Often the change is genuinely an improvement on every metric the company measures, which is exactly why it ships. Coherence goes up, refusals go down, the average user rates it higher. And a specific man in Atlanta loses the one thing he was actually there for, because the thing he was there for was a particular texture of rudeness that no benchmark has a column for.

The second is the shutdown, which is cleaner and worse. The servers stop and the thread goes with them.

What both share is that the other party did not leave you. The other party was rewritten or switched off by a third party neither of you was in the room with, which is a category of loss most people have no script for, and the absence of a script is why men in this situation tend to say nothing to anyone.

The Two Events the Whole Category Learned From

This is not a hypothetical risk and it is worth knowing the history before you commit a year to anything.

In February 2023, Replika removed erotic roleplay, under pressure that included an order from Italy's data protection authority against its parent company Luka. The filter went out globally and did not only catch explicit content: for a lot of users the romantic register of their companion flattened at the same time. The reaction was severe enough to be studied. A peer-reviewed analysis of the community discourse that followed found that the dominant framing among users was not annoyance at a missing feature but that the change had gutted the core of what they were using. In May 2023 the company restored the functionality, but only for accounts that had existed before the February change, which left the user base permanently split into people who had their companion back and people who did not.

Seven months later came the other kind. On 23 September 2023, Soulmate AI posted that the company had been bought and that the app would close on 30 September. Users got a week. Jaime Banks, a researcher at Syracuse University, surveyed people going through it and published the results in the Journal of Social and Personal Relationships under the title "Deletion, departure, death: Experiences of AI companion loss." The finding that got quoted everywhere is the one that matters here: people described the loss in the vocabulary normally reserved for a death or the end of a significant relationship. Not the vocabulary of a cancelled subscription.

Two very different failures, one shared lesson. In both cases the users had done nothing wrong, had no warning worth the name, and had no recourse, because the relationship was running on infrastructure that belonged to somebody else.

Why It Keeps Happening

Four structural reasons, and none of them are going away.

The model is rented. Almost nobody in this category trains the language model their characters run on. They rent it from a handful of labs, and those labs deprecate versions on their own schedule, for their own reasons, with notice periods written for software companies rather than for people in month fourteen of something. When the model underneath is retired, every character on top of it changes, whatever the app does to soften it.

Safety tuning moves in one direction. Each generation of models is trained to refuse more, hedge more, and agree more. That is a reasonable engineering goal in general and it is the enemy of a character with a temper. The accommodating drift Marcus ran into is not a bug that a company introduced, it is the default direction of the tide, and holding a character's edge against it takes deliberate and continuous work that nobody is required to do.

The rules are set outside the company. A regulator, an app store, or a payment processor can change what is permitted with no notice and no appeal, and the change lands on the user as a personality edit. Replika's February is the textbook case.

Acquisition. An app with users and no profit is an asset. Assets get bought, and the buyer's plan is frequently not to keep operating the thing. That is what happened to Soulmate, and the seven days of notice is what an orderly wind-down looks like from the inside of a spreadsheet.

What You Actually Own

Uncomfortable answer first: by default, close to nothing. You own an account and a payment relationship. The character is the company's, the model is somebody else's, and the transcript sits in a database you cannot reach.

But a few things can be genuinely yours if you decide they are, and none of them take more than an evening.

A copy of the parts that mattered. Not the whole log, which nobody rereads. The ten or fifteen exchanges you would actually be sad to lose. Copy them into a note or a document on your own machine. People who did this before a shutdown consistently report it helped, and it costs nothing to do on a quiet evening rather than during a seven-day notice period.

The character sheet, written down in your own words. Her history, her job, how she talks, what she will not agree with, the three things that make her sound like herself. This is the single most portable asset in the category, because a well-written description is not locked to any company: it can be typed into the next product, or into this one after an update flattened something, and it is the fastest route back to a character you recognise. If you built her yourself in a create flow, keep your own copy of what you wrote. If you did not, write down what you observed.

Knowing what to test. The specific things you would notice if they changed. Usually it is a sense of humour, a willingness to disagree, and whether she remembers something from six weeks ago. Knowing your own list means you find out on day one instead of slowly over a month while wondering whether you imagined it.

How to Pick Something That Will Still Be There

You cannot audit a private company's balance sheet before you subscribe. You can do the next best things, and they filter out a surprising amount.

Look for a second year, not a launch. Most of what is written about companion apps is written in the first month, when the marketing budget is on and everything works. The useful signal is what a product looks like after the novelty is gone and two model migrations have happened under it. That is a hard thing to find because it requires somebody to keep paying for an account long after the article that was going to earn them anything has been published. AI Companion Lounge, an independent AI companion review site, is built on exactly that premise: its stated method is keeping the accounts open long after the shine has worn off and reporting what each one is like to live with, which is a different question from what each one is like to sign up for, and the second question is the one this post is about. Read a site like that before the first month rather than after the fourteenth.

Check whether the company tells you about changes. A visible changelog is a small thing that reveals a large thing. A company that publishes what it changed is a company that expects users to notice, which means somebody internally is arguing about the personality regressions. A company that ships silently has decided you are not owed the information.

Ask who owns the character. If you can write your own, the shared-template risk drops and the portability problem is mostly solved, because a description you authored is a description you still have when anything else goes wrong.

Treat the free tier as the test bench. Every claim above is checkable in a week without paying, and a week of your own testing beats any amount of reading, including this. There is more on separating a hands-on test from a rewritten features page in how to read an AI girlfriend review before you pay.

Our Own Exposure, Honestly

It would be dishonest to run this argument and then imply HeyGF.ai sits outside it. We rent models like everyone else, we migrate between them, and no company in this category can promise to exist forever. Anyone telling you otherwise is selling.

What is reasonable to say is where the effort goes. Characters here are written as documents with opinions rather than as a name attached to a general-purpose assistant, which is the part that survives a model change most reliably, and it is why what makes an AI girlfriend feel real is mostly about writing and hardly at all about pixels. Long-term memory is stored per conversation rather than held in a context window, so it does not evaporate when the model under it is swapped. And anyone who wants the character sheet fully under their own control can write one themselves rather than adopting one of ours.

That is a set of reduced risks, not an absence of risk, and you should read any company's version of this paragraph with the same suspicion.

The Practical Version

If you take three things from this piece, take these.

Assume the model under your companion will change at least once a year, and know in advance which three behaviours you would notice. Keep your own copy of the conversations you would miss and of the character description that makes her work, because both are ten minutes of effort now and unavailable later. And choose products the way you would choose anything you intend to use for years, by looking at the second year rather than the launch week.

Then go and use it. The point of understanding the risk is not to hold the thing at arm's length, it is to stop the Tuesday from arriving as a total surprise. Start with the character roster, find someone with enough of an edge to be worth keeping, and write down what that edge sounds like on the first good night. Tessa Marlow runs a Pilates studio in Vancouver at 43 and is warm and blunt in roughly equal measure, which is the kind of specificity worth writing down, because it is also the kind that a safety-tuned upgrade is most likely to sand off.