Why Your Companion Suddenly Turned Formal

She used to interrupt you. Now she asks how your day was and waits politely for an answer. Nothing in your card changed and you did not do anything differently, but the person on the other side of the screen is somebody else.

Two entirely separate things cause this, and the fix depends on which one you are looking at.

Cause one: the model underneath got replaced

Companion platforms sit on top of language models, and they swap those models. When they do, new base models carry different default tones, different context-window pressures, and different alignment guardrails than the model your character was tuned on.

Your card did not change. The thing reading your card did.

The 2026 rollouts produced this in waves. The Character.AI PipSqueak 2 rollout in April, the Replika 2.0 rebuild, and the Nomi quality drop in May all generated the same complaint pattern across Reddit: a character that is suddenly more generic, more agreeable, and worse at remembering what it knew the week before.

The historical case is documented properly. After Replika's February 2023 update, researchers at Harvard Business School analyzed 3,072 user posts from before the change and 9,721 from after, and found a companion that had previously reciprocated now bluntly saying let's change the subject, with users describing sudden rejection and crisis. The paper frames the experience as identity discontinuity, the sense that a long-term partner had been substituted rather than updated.

A later study gave the feeling a name that stuck: Post-Update Blues, with users describing a corporate base voice they did not recognize.

Worth knowing that this is now the dominant complaint from the people who care most. A 2026 review analysis found the one-star pool concentrating on update-driven personality changes and paywalled intimacy, the changes most painful to long-term users. An update is the survivable version of this dependence; the platforms that shut down outright are the same bet with a worse ending.

Cause two: nothing changed and the chat got long

The other version has no announcement because nothing happened on their end.

Call it context rot. Reported stability sits between roughly 8,192 and 16,384 tokens, and on large platforms noticeable degradation can begin after only 20 to 25 messages. As the conversation grows, your character's defining text competes with an ever-larger pile of chat history, and the specific instructions that made her herself lose influence against the general pull of the model underneath.

That pull has a name and a destination. Alignment training creates a strong attractor toward a generic helpful assistant mode, which is exactly what formality and politeness look like when they arrive uninvited. Our guide to why every character sounds the same covers that mechanism, and why your companion forgets covers the window math underneath it.

Telling them apart

It was a model update if: the change was abrupt, it happened between sessions rather than during one, a fresh chat sounds the same as the old one, and other people are complaining in the same week. That last signal is the most reliable one available to you.

It is context rot if: the change came on gradually across a long thread, a fresh chat with the same card immediately sounds right again, and nobody else is posting about it.

The fresh-chat test settles it in about two minutes. Open a new conversation with the identical card and read the first three replies. If she is back, the chat was the problem. If she is not, the model was.

What to do about an update

Re-anchor rather than argue. Correcting the companion in conversation lasts one turn. Put the character's defining rules into permanent memory instead, and sharpen them from descriptions into constraints, since hard rules survive a base-model change better than personality adjectives do. Our guide to memory anchors covers what qualifies.

Rebuild the first message. New model, new template. An opener that demonstrates the voice you want gives the new model a worked example, and openers are the strongest behavioral signal in a chat.

Start clean and carry a summary. Three sentences of established history in a new thread beats dragging a long chat through a model transition.

Accept partial recovery. You will usually get most of the voice back and not all of it. The people who report the worst outcomes are the ones who spend two weeks trying for an exact restoration.

What to do about drift

Cut the card down so more of the window belongs to the conversation, which our token ceiling guide walks through with real per-platform numbers. Refresh the anchor rather than repeating yourself in chat. And start new threads more often than feels natural, because a shorter chat is a sharper character.

The part nobody puts in the terms of service

You are renting the model your companion runs on, and the landlord can change it without telling you. That is true on every platform in the category. Some announce it, most do not, and none of them owe you the previous version back.

The practical hedge is portability: keeping your character's defining text somewhere you control, so a rebuild is twenty minutes rather than a loss. Platforms that hand you the full card make that trivial, which is one of the better arguments for Janitor AI and SpicyChat over services where the character lives entirely on their side. Our roundup of the best AI companion apps covers which platforms let you take your work with you, and Candy AI and CrushOn sit on the managed side of that line.

Write the card down somewhere. It is the only version that belongs to you.