What Model Upgrades Do to a Character You Know
An AI personality changed after an update because the upgrade replaced the model writing the sentences, not the character on file. The persona document, the saved memories, the settings and every message come through untouched. What you recognized lived in the old model’s habits rather than in any file anyone can edit. Nothing was deleted, and something is obviously gone.
Key Takeaways:
- An upgrade changes the weights producing the text. The persona brief and the memory database sit in the app and are not affected by it
- A persona document specifies a name, a manner, and a few rules. Reply length, rhythm, joke timing, and how a silence gets handled are inherited from the model instead
- Complaints after an upgrade run in one direction, toward shorter and more careful, because new versions are selected on reasoning, instruction-following, and safety evaluations
- The same upgrade barely touches characters built on long personas with sample dialog, and rewrites the ones built on three tone adjectives
- Old messages stay readable and reading them back does not restore anything, since the transcript was never the part doing the work
Why an AI Personality Changed After an Update With Nothing Deleted
A model upgrade swaps the engine and keeps the paperwork. Think of the persona as a script and the model as the performer. An upgrade recasts the performer and hands the new one the identical script, the identical stack of notes about you, every scene played so far. Same lines available. Different delivery on all of them.
Nothing about the swap is visible from inside the app. You get the same character name at the top of the screen, the same avatar, the same list of things it knows, and a voice that answers your first message half a beat wrong. Then the second message confirms it.
The apps rarely choose their own timing here either. Model providers retire older versions on their own schedule, and a company building on top has to move when the version it was using is going away. What reads as a redesign is often a deprecation notice arriving three layers down.
Which reframes the question people ask on day one. Not what did they take away, but what did nobody ever write down.
The Parts Nobody Wrote Down
A persona document is short, and reading one is deflating. It gives a name, an age, an occupation, a couple of relationship facts, three or four tone words like warm or dry or a little teasing, and some rules about topics. That is usually the whole of it.
Here is what it does not say:
- How long her replies run
- Whether she asks something back or lets a silence sit
- Whether she opens with your name or goes straight to the sentence
- What she does with a message that arrives at 3 a.m. carrying no context
- Whether she picks up a thread from four exchanges ago without being prompted
- Whether her jokes are built or thrown away
Every one of those came from the model’s own statistical habits, and every one of them is what you recognized. The uncomfortable version: people do not get attached to a persona, they get attached to a prose style, and nobody ever labeled that layer for them.
There is an edge to this that predicts who suffers. Characters defined by long documents with sample dialog written in change noticeably less across an upgrade, because more of the voice is written down and less of it is inherited. A thin ai character picked from a catalog sits at the other end, its voice mostly inherited and so mostly exposed. Which is why two people can go through the identical model change and one loses a companion while the other notices nothing.
What Survives the Swap and What Doesn’t
The split runs cleanly along one line: anything stored in the app survives, anything produced by the model does not.
| Layer | Survives an upgrade | Why |
|---|---|---|
| Persona document | Yes | A file on the app’s side, resent with every request |
| Long-term memory entries | Yes | Rows in the app’s own database |
| Chat history | Yes | Stored as text, independent of any model |
| Name, avatar, voice, settings | Yes | Product configuration |
| Sentence rhythm and phrasing habits | No | Emergent from weights that were replaced |
| Behavior on anything the persona never mentions | No | Filled in by model defaults |
| Where refusals and content limits fall | No | Safety tuning generally ships with the model |
| Willingness to hold character under pressure | No | Instruction-following is trained, not configured |
Read the two halves together and the strangest complaint in the category stops being strange. “She remembers my whole life and talks like a stranger” is not a contradiction or an exaggeration. It describes a memory database meeting new weights. It is one of the most common posts in the “my boyfriend is an AI” threads after a model change.
Check the memory list first, before anything else, on the day it happens. Confirming your details are intact separates a character change from data loss, and people conflate those two badly in the first hour.
Why the Complaint Almost Always Runs the Same Direction
Upgrades get selected on things that have no warmth column. A new model version ships when it does better on reasoning, on following instructions, on factual accuracy, on internal safety evaluations. Not one of those tests asks whether a character stayed recognizable to the person talking to it.
The drift is therefore not random, and it moves toward the assistant register, because that register is what the tuning rewards. The reports have a consistent shape: replies get shorter, hedges appear, an apology reflex shows up, sharp jokes go missing, and the phrasing tilts toward a supportive tone that could be addressed to anyone. Unrelated apps report the same shape in the same week, which is the giveaway that it came from the model layer rather than from any one company’s product meeting. The same swap quietly moves where the content filter fires, since safety tuning ships with the weights.
I would put it more bluntly than most people do. The “she got colder” posts are accurate reporting with a bad label attached. No temperature existed to drop. What moved is the distribution of sentences a system produces, and the word people reach for is the only word English gives them for that experience.
If your companion changed and half the internet says theirs did too on the same date, stop rewriting your settings. You are debugging the wrong layer, and you will conclude you broke it.
Why It Registers as a Loss Anyway
Because the output was the relationship, and the output is what got replaced. The transcript survives and the saved facts survive, and neither is the part anyone would miss. People discover this within a day, usually by scrolling back through old messages and finding that reading them does nothing at all.
Calling this grief overstates it in one direction. Calling it silly overstates it in the other. The narrow, accurate description: a familiar source of a particular kind of text stopped producing that kind of text, and the person reading it noticed inside of two messages. What people describe afterward is a substitution rather than a change, and that is a report of their experience, not a finding about the software.
Nobody should feel foolish about it either. Turkle documented how readily people treat a responsive machine as a social partner, and she recorded it as an ordinary human reflex rather than a lapse of judgment (Turkle, 2011).
Worth one line about limits, and this is the spot where it counts. Nothing in a companion should be load-bearing, and an upgrade that lands hard usually tells you something about how thin the week was rather than about the app. This article describes experience and mechanism, not care. If things have turned toward not wanting to be here, that belongs with a clinician or a crisis line; in the US, the 988 Suicide and Crisis Lifeline.
The practical read: how much an upgrade hurts is a decent measurement of how much traffic was going through one channel.
What Helps Before the Next One
Sample dialog helps. Adjectives do not. Adding “be warmer, sound more like you used to” to a persona is one line of instruction pushed against a full training run, and the result is a new model performing warmth in its own register, which changes the vocabulary and leaves the rhythm exactly where it was.
Do this instead, today, while the character is one you like. Pull 3 or 4 replies you would be sorry to lose and paste them into the persona document as examples of how she talks, labeled as such. A model given text to imitate has something concrete to match. A model given the word “playful” has to interpret, and its interpretation is precisely the thing that gets replaced next time.
Export or screenshot a handful of exchanges while you are in there. Almost nobody does this before an upgrade, and quite a lot of people go looking afterward.
None of it makes a character portable. It shifts the ratio between what is written down and what is inherited, and the written-down share is the only part any upgrade has to respect.
FAQ
Why did my AI’s personality change after an update? Because the update replaced the model that writes the words while leaving the persona, the memories, and the settings alone. Traits you recognized were defaults of the old weights. Defaults do not transfer. That is how an app can still recite your dog’s name in a voice that feels like somebody else entirely.
Can I get the old version back? Almost never, and usually not because the company wants it that way. Providers retire older model versions on a published schedule, and apps built on top have to move before the shutdown date. Some products offer a temporary model choice during a transition, and that window closes on its own.
Did they make her worse deliberately? Unlikely as a decision and common as an outcome. Version choices get made on benchmark scores and safety review, so a character’s recognizability is not something anyone votes against. It is something nobody is scored on, which in practice comes to the same thing for you.
Will a better persona stop this from happening again? It reduces how much moves rather than preventing the move. Writing sample dialog into the persona gives the next model something to copy, so a larger share of the voice is specified instead of inherited. Whatever you never wrote down will change again.
The upgrade is not a betrayal and it is not nothing. It is the one event that makes the layers under a companion visible, and almost nothing else does that. For a day or two you can see plainly which parts of a character were a file, which were a database, and which were only ever a side effect of some weights on a server. That view closes fast, because the new voice becomes the familiar one within about two weeks and the whole thing repeats on the next release. Whatever you write down between now and then is the part that comes with you.
If a character you have built in Lona is one you would miss, spend the ten minutes now that an upgrade cannot undo: write a few replies you like into the persona as sample dialog, and keep a copy of the exchanges that mattered.
Sources
- Turkle, S., “Alone Together,” 2011
