You build a character, spend an hour getting them specific, and somewhere around the fiftieth message they aren't them anymore. Same name, same avatar, different person. An AI chatbot changing personality partway through a conversation is one of the most common complaints in companion chat, and almost everything written about it recycles the same handful of explanations.
Most of those explanations are wrong, or right for a reason nobody states correctly. Which is why the fixes people pass around have such a poor hit rate.
Here are six beliefs worth dropping, and what's actually happening underneath each one.
Myth 1: it forgot who it was
The forgetting frame feels right because that's how people fail. It isn't how the model fails.
On most platforms, the character description doesn't go anywhere. It gets sent again with every single message, sitting at the top of the request like a header. Nothing deletes it. The bot that just answered you in a flat customer-service voice was still holding a description that says your character is a sarcastic ex-soldier who hates being asked questions.
What changed is the competition. The model isn't reading the description and then deciding whether to obey. It's producing the most likely next line given everything in front of it, and by message fifty, most of what's in front of it is transcript.
There's no internal switch labelled "in character" that flips off. There's a prediction that either resembles your character or doesn't, and the resemblance is a matter of degree. If you want the longer version of what happens between your message and the reply, we walked through the whole pipeline separately.
The distinction isn't academic. If it forgot, you'd remind it. Since it didn't forget, reminding is one of the weaker moves on the list, and it's the move everybody reaches for first.
Myth 2: drift only goes one way
Nearly every guide describes drift as a slide toward blandness. That's one of three directions, and which one you're looking at is the most useful piece of information you have.
Flattening
The character loses edges. Replies get longer and much more agreeable. Someone who used to give you four words now gives you a paragraph with a summary at the end.
This is the default-assistant pull. It shows up when your character's specifics have thinned out against a very large body of ordinary helpful text, and it's most common in long conversations that have gone quiet on plot. Nothing in the recent transcript demanded a strong reaction, so nothing strong came back.
Sharpening
The opposite failure, and almost nobody names it.
Instead of fading, one trait gets louder every message until the character is a parody of themselves. The aloof one turns cruel. The cheerful one turns manic. The flirt turns into something you didn't ask for.
This isn't the model losing the character. It's the model following recent evidence exactly. You responded warmly to a sharp line, the sharp line showed up again, and two instances of a pattern are enough to make the third likelier. Sharpening is a feedback loop, and loops accelerate.
Assimilation
The character starts sounding like you. Your sentence length, your punctuation habits. Type in lowercase with no periods for long enough and so will they.
This one hides well, because the character still says character-appropriate things. It's the register that's quietly become yours. Assimilation is worst when you're the only other voice in the scene and your messages run longer than theirs. It's also the direction most likely to end with the bot writing your side of the conversation, which is a related failure with its own set of fixes.
Three directions, three different causes. A fix aimed at flattening does nothing for sharpening. That's most of why generic advice misses.
Myth 3: a longer, more detailed description would fix it
Past a point, more description makes drift worse. Worth understanding why before you go back and add another five hundred words.
A character description isn't a rulebook. It's evidence about what this person is like, and the model samples from it. Every trait you add is one more thing that can get sampled on any given turn.
So a description full of traits that pull against each other will read as inconsistent no matter how well the model behaves. "Cold on the outside, warm underneath" is a fine line for a novelist, because a novelist decides when each side surfaces. Hand it to a text predictor and you get cold on Monday, warm on Tuesday, nothing motivating the switch. That reads exactly like personality drift, even though the model is following your description faithfully.
Adjective lists have the same problem. Ten traits with no relationship to each other produce a character who is one of those ten things per message, roughly at random.
What works better is fewer traits with conditions attached. Not "sarcastic and caring" but "sarcastic by default, drops it when someone's actually in trouble." Now the switch has a trigger the model can detect in the transcript instead of picking a side each turn.
This is also why pre-written characters often hold shape better than ones people build in a hurry. Not because someone wrote more, but because the traits were tested against each other first.
Myth 4: the content filter did it
Filters do interrupt characters, and on some platforms they interrupt them a lot. But the filter takes the blame for drift it had nothing to do with, and the two look different once you know what to look at.
A filter intervention is abrupt. One message is in character, the next has swerved, and there's often a sentence in there that sounds like it came from a different document: a redirect, a softened refusal, a suggestion that you talk about something else. No gradient. It happens between two consecutive turns and it usually attaches to a specific topic.
Drift is continuous. Scroll back and you can find the slope. Message thirty is a little softer than message twenty, forty is softer still, and there's no single turn where it broke.
So: if you can point at the exact message where it changed, look at what you sent right before it. If you can't find that message, the filter probably isn't your problem. Filters are also the first thing people compare when they're picking a platform, and close to the least useful thing to compare, for roughly this reason.
Myth 5: regenerating will bring them back
Regenerating resamples from the same conditioning. Same description, same transcript, same recent turns that pulled the character sideways in the first place. You'll get a different sentence out of the same distribution that produced the one you didn't like.
Sometimes that's enough. The distribution is wide and you got unlucky.
Often it isn't, and here's the part that costs people: every regenerated reply you accept adds another off-character message to the transcript, which makes the next one likelier to be off too. Reroll five times and keep the best of five, and you've still added an off-character message to the evidence pile.
Editing beats regenerating, for the same reason it does with repetition. We laid the mechanics out in the piece on breaking a chatbot out of a loop. Short version: the transcript is the evidence, so change the evidence.
Myth 6: nothing you do matters, it's the model
The fatalistic version, and the expensive one, because it stops people trying the things that work.
Model quality sets a ceiling. It doesn't set the floor, and most conversations sit nowhere near the ceiling. Character consistency responds to scene design far more than it responds to prompt wording, and scene design is entirely yours.
What actually holds a character in place
Give them something to want. A character with a goal in the current scene has a reason to stay consistent, because consistency is now useful to them. A character with no goal is a voice, and voices drift.
Put something else in the room. Assimilation and flattening both feed on two-person conversations with no external pressure. A third character, a deadline, a locked door, weather, anything that isn't you gives the model something to react to besides your last message.
Let them disagree with you. If every message you send is met with agreement, you've trained a mirror. A lot of drift is downstream of a chat where nothing was ever contested.
Move the scene when it goes quiet. Flattening loves a conversation that's run out of situation. Changing location resets what the recent transcript is made of, which is the thing doing most of the voting.
Then check the direction before you pick a fix. Flattening wants stronger situations. Sharpening wants you to stop rewarding the loud version, and often to go back and edit the message where it first got loud. Assimilation wants you to write shorter and let them write longer. Same symptom, three different jobs.
If none of that lands, starting fresh is underrated. Not because the old chat was ruined, but because a clean transcript is the one condition you can change completely, and it takes a minute. People treat a long chat as an investment and keep nursing it. Some of them are worth nursing. Many aren't.
Where Friend2Chat sits on this
Straight answer first: not solved here either. Friend2chat runs on the same class of model as the rest of the category. There's no persona builder, no scene editor, no way to pin a trait so it can't be outvoted by the transcript, and long conversations drift here the way they drift everywhere else. Anyone telling you their app fixed this is describing a roadmap.
What's different is narrower. Characters come pre-written with traits already tuned against each other instead of listed, which removes the most common self-inflicted version of myth three. The interface shows what the character currently knows about you, so when the register shifts you can check whether you're looking at assimilation or something else instead of guessing. And you can open a character and start a scene without filling in a form, which matters more than it sounds when the fix you're testing is "start over somewhere new."
That last one is the honest pitch. If you want an online AI to talk to while you're working out which direction your character is drifting, the useful property isn't a better memory system. It's that a fresh scene costs you nothing, so you can run the test five times in an evening.
If you're newer to this and half the terms above were unfamiliar, the common beginner mistakes piece covers the ground underneath this one, and the roleplay setup guide covers scene design in more depth than a myth list can.
FAQ
Why does my AI character get nicer over time?
That's flattening, the default-assistant pull. It shows up most in chats that have gone quiet on plot. Introduce a situation the character has to react to, or move the scene somewhere new.
Can I stop a character from copying my writing style?
Partly. Write shorter messages than theirs, don't answer in their register, and put a third voice in the scene. Assimilation is strongest in a two-person exchange where you're doing most of the talking.
Should I edit the bot's message or regenerate it?
Edit. Regenerating draws from the same conditioning, while editing changes the transcript the next reply gets built on.
Does drift mean the character was written badly?
Usually not. More often the description holds traits that pull in opposite directions with no stated trigger for which one surfaces, so the model picks one per turn and you read the result as mood swings.
Do I need to sign up anywhere to test this?
Not everywhere. On some platforms you can speak to an AI online and start a scene straight away, which is enough to see how a character holds up over thirty or forty messages. If you want to talk to an AI for free while you compare a few, that trial run is the whole test.
