Of all the complaints people make after months with a companion app, this is the big one, and it gets misdiagnosed constantly. The usual theory is that the app lost the character. In most cases she is right where you left her, just outnumbered by everything you have typed since.
Four reasons characters change
Three of them creep up during chats. One can turn up in a day.
Fading
Think of each reply as being written by a model that reads her profile and then the latest slab of chat. Early on the profile is most of the page. Months in, it is a sliver, and the chat is nearly all the model has to go on. So she gradually becomes whatever the recent messages hint at.
Clue: the change is slow, and a brand-new thread makes her normal again.
Echoing
Language models pick up your rhythm. Write brief, sweet messages and hers turn brief and sweet. Have a gloomy week and she stays gentle far longer than you meant. Since these systems also lean towards agreeing, a strong-minded character can end up as a mirror.
Clue: she reads more like you with every passing week.
Ruts
A turn of phrase, a gesture or a pet name that has come up a few times gets likelier each time it is used. A while later it is in every message. The rut deepens itself, because each use counts as one more sample to imitate.
Clue: you can predict her sign-off before you read it.
Model swaps
Companies change the model underneath, along with safety rules and how memory works. A different model handed the same profile speaks differently. This hit a lot of Replika users in 2023, and we cover it in when your AI companion changes overnight.
Clue: many users notice on the same day, usually around an update or announcement.
Which repair for which cause
| If it is | Try this |
|---|---|
| Fading | Start a new thread every so often, having saved the essentials to memory |
| Echoing | Write her key traits into the profile as fixed rules, and change up your own style |
| Ruts | Rewrite or reroll the first few repeats, and put a line in the profile asking her to vary her wording |
| A model swap | Redo the profile with the new model in mind and add samples of her speaking |
Samples beat adjectives
If you do only one thing, add two or three lines she would actually say to the profile or the example-dialogue box. A model copies a sample much more faithfully than it obeys a description. The word "sarcastic" gives it little to work with, but "Brilliant. Another master plan. I'll pack the first-aid kit." gives it plenty.
Keep a copy of those lines somewhere outside the app. After a model swap, or if you ever change services, they let you rebuild her voice. More on that in switching AI companion apps.
A once-a-month check-up
- Reread the profile and your sample lines.
- Go through her last twenty replies and ask whether they still match the samples.
- If they do not, decide which of the four reasons fits and use its repair.
- If she has wandered somewhere you like better, rewrite the profile to match. Drift is only a problem when you did not choose it.
The app is not failing here. A system built to continue patterns will behave this way in a long chat. A little upkeep keeps her the character you designed, and skipping it tends to lead to the flat, samey feeling covered in when the novelty fades.