What Survives a Model Swap and What Does Not

A model swap does not delete anything. Every word of your card is still there, every message of your history is still stored, and the character reads completely differently anyway.

That is because the swap changed the reader rather than the text. Here is the audit, field by field, of what actually carries.

The survival table

| What you built | Survives? | Why | |---|---|---| | Card text as stored data | Fully | It is a database field, untouched by a model change | | Chat history | Fully | Stored separately from the model | | Hard behavioral rules | Mostly | Constraints are checkable regardless of model | | Refusals | Mostly | The most model-independent thing you can write | | Speech signature | Partially | Depends on how strongly the new model's defaults pull | | Personality adjectives | Poorly | Absorbed into whatever the new model considers typical | | Prose style and rhythm | Rarely | This is the model, not the card | | Relationship tone in a long thread | Rarely | Context built for the old model reads differently to the new one | | Memory depth | Varies | New models bring different context-window behavior |

Survival table v1.0 — verified 25 Jul 2026, rechecked quarterly as new model rollouts land.

The pattern in that table is worth stating outright: the more your character depended on the model's own qualities, the less of it survives. The more your character depended on explicit rules you wrote, the more of it does.

Why rules outlive adjectives

New base models carry different default tones, different context-window pressures, and different alignment guardrails than the model your character was tuned on. A personality adjective is an invitation for the model to interpret. A new model interprets it differently, using its own defaults, and its defaults are the thing that changed.

A behavioral constraint is not an invitation. "She never reassures anyone" is a rule any model can check its output against. It is the least model-dependent instruction available, which is exactly why it survives transitions that flatten everything around it.

There is a deeper reason. Alignment training pulls every model toward a generic helpful attractor, covered in our guide to why every character sounds the same. Descriptions get absorbed into that attractor. Refusals resist it, because the attractor contains no refusals.

The pre-swap checklist

Run this before the next update rather than after, since the next one arrives without notice.

Export the card if the platform allows it. Our portability table shows which do. On the platforms that do not, copy the text into a file you control.

Convert your best adjectives into rules. Go through the personality field and find every trait that describes rather than constrains. Rewrite the two or three most important ones as behaviors. This is the single highest-value pre-swap edit and it improves the character immediately regardless.

Write down one hard refusal. If your card does not have one, it is more fragile than it needs to be.

Save the opening message separately. It is the strongest behavioral template in any chat and the first thing worth rebuilding after a transition. Our guide to what belongs in a character card covers why it carries so much weight.

Record what the character sounds like. Three or four lines of dialogue that felt exactly right. After a swap, memory of the old voice gets unreliable fast, and a sample gives you something concrete to aim at.

The post-swap recovery order

Diagnose first. Open a fresh chat with the same card. If it sounds right, this was chat length rather than a model change, and our guide to why your companion turned formal covers the difference.

Re-anchor rather than argue. Correcting the companion in conversation lasts one turn before it scrolls away. Persistent instructions are a different mechanism, covered in our memory anchor guide.

Rebuild the opener. New model, new worked example. An opening message that demonstrates the voice you want does more than any amount of description.

Start clean, carry a summary. Three sentences of established history in a fresh thread beats dragging a long chat through a transition. The old thread's context was built for a model that no longer exists.

Stop at good enough. Most people recover most of the voice. The people who report the worst experience are the ones who spend two weeks chasing an exact restoration that is not available.

The structural fix

Everything above is damage control. The actual fix is not being subject to somebody else's model decisions.

Bring-your-own-key setups through Janitor AI and self-hosted rigs via SillyTavern put the model choice in your hands, which means a swap only happens when you decide to make one. That is a genuinely different relationship to the product, and it is a better argument for the harder path than the filter argument usually made for it.

For everyone staying on managed platforms, and most people should, the hedge is portability plus rules. Candy AI, CrushOn, and SpicyChat all deliver more polish than a self-assembled setup, covered in our best AI companion apps roundup. Just keep your character's defining text somewhere they cannot reach it, and check the model change log when something feels off.

The card belongs to you. The model never did.