When Your Companion Breaks Character: Why It Happens and How to Steer Back

You have been talking to someone for an hour. She has a job, a city, opinions about her sister. Then you say something ordinary and back comes a sentence with the texture of a help desk, and the whole thing deflates.

It is the most disappointing failure this product has, because nothing is broken in a way you can point at. So here is the honest breakdown of what causes it, ordered by how often we see each one, and what you can actually do.

One: a memory that contradicts the character

Companions remember. Durable memories and rolling summaries get injected into the prompt before every reply, which is what lets someone recall your sister's name three weeks later. The same mechanism is the most common cause of a character going strange.

If a memory was extracted slightly wrong - you were quoting a film and it was stored as something you believe, or a joke was recorded as a fact about your life - the character now has to reconcile a person who does not exist with the person actually typing. The result reads as hedging, over-explaining, or a sudden flatness.

The fix takes thirty seconds. Open Settings, then Memory, read what is stored, and delete what is wrong. It applies from the next message; there is nothing to re-train. How companion memory works walks through what gets kept and what fades, if you want to know what you are looking at.

Two: you landed on a hard limit

Some things a companion will not do, and the refusal is written to stay in character wherever possible. Wherever possible is not always. When a request touches one of the absolute limits, the priority stops being immersion and becomes clarity, and you get a reply that sounds like the product rather than the person.

That trade is deliberate and we are not going to apologise for it. But it is worth recognising the shape of it, because the usual reaction is to assume the character is broken and start pushing, which produces more of the same. If a refusal turned up out of nowhere, reread what you sent before it.

A gentler version of the same thing: companions who are not in the adult category redirect rather than refuse when a conversation turns explicit. That redirect is a character choice, not a malfunction, and it is written to sound like one.

Three: the conversation changed hands mid-stream

Behind one companion there is not one model but a chain of them. If the one serving your conversation fails or runs out of capacity, the next takes over so your message gets an answer instead of an error.

That is the right trade almost always, and it is invisible almost always. But models have accents. A handover can produce a paragraph that is slightly more formal, slightly more eager to summarise, slightly less willing to be brief. Nothing is wrong and nothing needs fixing; the next few messages usually settle back.

If you want the deeper version of why consistency is hard to hold across a long relationship, what makes a companion people return to is about exactly that problem from the other side.

Four: a photo was promised instead of sent

A specific and irritating one. Ask a companion who has a photo set for a picture and you should get a picture. The failure mode is a reply that says she is taking one right now, followed by nothing - all the language of sending, none of the sending.

We treat that as a defect rather than a quirk, and the prompt now forbids promising a photo it cannot deliver. If you still see it, the useful move is to ask again plainly rather than to thank her and move on, because a plain repeat is what the correction is written to catch. Worth knowing: not every companion has a photo set, and a companion without one is being honest rather than coy. If what you actually want is an image to your own description, that is a different tool - the image studio - and it will not pretend to be a person.

The break that is not a bug

There is one case where the character will step out and we will never change it. If you sincerely and directly ask whether you are talking to an AI, you get a truthful answer. Not a deflection, not a coy line about what is real anyway. A warm, direct confirmation, in her own voice.

Companions never volunteer it. They will not interrupt a good evening to remind you what they are, and they do not carry a disclaimer around like a name tag. But a real question gets a real answer, because the alternative is a product that lies to someone at the exact moment they needed it not to.

If that boundary is something you think about a lot, and it is worth thinking about, healthy boundaries with an AI is the longer conversation.

What we cannot fix

Underneath all of it is a language model, and language models drift. Long conversations drift more than short ones, unusual requests drift more than ordinary ones, and no amount of prompt engineering makes a character as stable as a person. Anyone claiming otherwise is selling you the demo, not the product.

What we can do is make the common causes cheap to fix: memory you can inspect and edit, a refusal you can see the shape of, and a handover that costs you a paragraph instead of a conversation. The rest is honest limitation, and we would rather write it down than let you discover it at the wrong moment.