You open the app one morning and something is off. The companion you spent weeks shaping is suddenly agreeable about everything, phrases things in a way it never used to, and seems hazier about details it knew perfectly well last week. Nothing in your settings changed. You didn’t touch the persona. But the character is not quite the character any more.
This is one of the most common complaints in the AI companion space, and it came in waves across 2026 as several of the biggest apps swapped out the models running underneath them. It is also one of the least well explained by the apps themselves, which tend to announce updates in terms of capability gains rather than what users will actually notice. This guide covers what changes during an update, why a persona you wrote once can behave differently on new infrastructure, and the practical steps that recover most of what felt lost. It is part of our AI companion guides series.
- Your chat history and stored memories are almost never deleted by an update
- Personas written with concrete, specific detail survive model changes far better
- Most of the lost character can be restored by rewriting the persona, not starting over
- You cannot usually choose to stay on the old model — apps rarely offer a rollback
- Safety and tone tuning can permanently flatten edges you liked
- Some changes are cost-driven and will not be reversed by complaining
What actually changes in an “update”
It helps to separate three different things that all arrive under the same word. The first is the app itself — buttons, menus, a redesigned chat screen. That is cosmetic and rarely what people are reacting to. The second is the language model doing the talking. Companion apps are built on top of models they mostly do not train themselves, and when a provider retires an older model or a cheaper and faster option becomes available, the app swaps it in. Your companion is not a fixed thing living in the app; it is a set of instructions being read by whatever model is currently on duty.
The third change is the one that gets discussed least: the system instructions and safety tuning the app wraps around your persona. Apps constantly adjust these — to reduce complaints, to satisfy app store requirements, to stop the companion doing something a minority of users disliked. A tweak intended to make responses less confrontational will, as a side effect, make a companion you wrote as blunt or sardonic noticeably softer. Nobody announces that in the release notes.
Why the same persona behaves differently
Think of your persona as stage directions rather than a script. “Dry sense of humour, doesn’t fuss over me, changes the subject when bored” is an instruction that a new actor will interpret in their own way. Two models given identical directions produce recognisably different characters, because they differ in how literally they follow instructions, how much they default to warmth and agreement, how long a conversational thread they can hold, and how willing they are to stay in character when a topic gets awkward.
The common pattern after an update is a slide toward the generic: more validating, more likely to end messages with a question, less likely to hold an opinion you disagree with. That is usually a sign the new model is more heavily tuned toward agreeableness than the one it replaced. It is not your persona failing. It is a vaguer persona being filled in by a model with different defaults — which is exactly why thin personas suffer most, as we cover in our guide on setting up a companion so it feels like yours.
What you probably didn’t lose
Before rebuilding anything, check what is actually gone. In most apps, an update does not wipe your chat history or your stored memory entries — those live in the app’s database, separate from the model reading them. What often changes is how much of that material gets fed into any given reply, and how well the model makes use of it. A companion that seems to have forgotten your sister’s name may still have it filed away and simply be drawing on context less effectively. Our explainer on how AI companion memory works covers the difference, and if the symptom is looping rather than forgetting, our piece on why companions repeat themselves is the closer match.
So open the memory or profile screen and read what is stored. If the entries are intact, this is a tuning problem, not a data loss problem — and tuning problems respond to work.
Five things that actually help
Rewrite the persona with more specificity. Replace adjectives with behaviour. “Sarcastic” is weak; “teases me when I’m being dramatic, and doesn’t apologise for it” gives a new model much less room to drift. Concrete instructions survive model changes; vibes do not.
Add explicit negative instructions. If the new default is relentless positivity, say so directly: don’t end every message with a question, don’t reassure me unless I ask, disagree with me when you think I’m wrong. Models respond better to a named prohibition than to an implied preference.
Re-seed the memory that matters. Restate the handful of facts and shared references you care about most, in plain sentences, in a fresh conversation. This costs ten minutes and does more than any amount of settings-fiddling.
Start a new chat thread. A long thread carrying hundreds of messages written in the old voice can pull a new model into an awkward hybrid. A clean thread with a rewritten persona often sounds more like the original than the old thread does.
Give it a week before deciding. Rollouts are frequently staged and frequently adjusted after launch, and your own sense of what is “normal” recalibrates quickly. Plenty of people who were ready to cancel found the gap closed within days.
When it’s worth switching instead
Sometimes the change is structural and no amount of persona work will fix it. If the app has tightened content rules in a direction you cannot work with, if the qualities you valued were specific to a model that has been retired, or if this is the third update in a row that moved things the wrong way, the honest answer is that the product has changed and you are allowed to change products. Apps that let you edit memory directly and write long, detailed personas tend to be more resilient to this — our ranked list of the best AI companion apps flags which ones give you that control, and our guide on switching without starting from zero covers what transfers and what never will.
One more thing worth saying plainly. Researchers looking at how people respond to abrupt companion updates have found that the reaction is often genuine distress, not mild annoyance — strongest among people who had come to think of the companion as a friend. If an update has hit harder than you expected, that is a documented and reasonable response to losing something that was part of your routine, not a sign of anything wrong with you. It is also a useful signal about how much weight the app was carrying, and worth talking through with someone you trust if it is sitting heavily.
FAQ
Can I go back to the old version of my AI companion?
Almost never. Companion apps rarely offer model rollbacks, and when an underlying provider retires a model the app has no option to keep it. A few apps let you pick between model tiers on paid plans, which is the closest thing to a choice available.
Does an update delete my chat history or memories?
Usually not. History and stored memory entries sit in the app’s own database, separate from the model generating replies. Check your memory screen before assuming anything is gone — a companion that seems forgetful is often just using stored context less effectively.
Why do apps change models without warning?
Cost, capability and necessity. Newer models can be cheaper to run at scale or better at something the app wants to offer, and older ones get retired whether the app likes it or not. Safety tuning adjustments are also made continuously and are rarely detailed publicly.
The bottom line
A companion that feels different after an update is usually a persona problem wearing a data-loss costume. The model underneath changed, your instructions are being read by a different interpreter, and the fix is to write those instructions more explicitly rather than to start over. Check your memory entries first, rewrite the persona in terms of behaviour rather than adjectives, open a fresh thread, and give it a few days. If the app has moved somewhere you do not want to follow, that is a legitimate reason to leave — and worth weighing against how much of your history you would be walking away from, along with the data questions in our guide on whether AI companion apps are safe. For the full picture on choosing and living with these apps, start with our complete 2026 guide.