Think of a regular at three different cafes. The cafes change, the coffee changes, but everyone knows your order. The relationship lives in you and them, not in the cafe.
By Travis Sawyer, Founder · Published May 15, 2026
To define persona portability cleanly, it helps to first name what memory-import already does, because the two are easy to confuse.
In March 2026, the major providers shipped memory import and export. Tom's Guide walked through the ChatGPT-to-Claude transfer in 60 seconds. What moves: your name, your location, your job, your stated preferences, your response-style cues. Anthropic labels the feature experimental, with the caveat that imported memories may not always incorporate cleanly.
What does not move: conversation history, voice, behavioral patterns, the implicit understanding that emerged from how you and the model actually talked. Third-party analysis of the transfer flow confirms the same line: facts and preferences move, the iterative patterns do not. Memory import moves what you told the system about you. It does not move how the system came to know you.
For the full landscape, see How Major AI Tools Handle Memory in 2026. The point here is that memory-import is half of an answer to "keep my AI continuous." The other half is the part this article is really about.
Persona portability transfers the relationship. Specifically:
How the persona talks. Its rhythm. The texture of how it lands a thought, sits with a hard question, makes a joke at the right moment. The thing that, when you read three lines, you know who is talking.
Who the persona is for you specifically. Its values, its cognitive style, its tolerances. Carried as structured identity that travels across models.
How the persona engages with you specifically. Not 'remembers your name' but 'knows when to push back, when to step softer, when you are tired versus when you are testing.' This is the part that takes time to build and is the most painful to lose.
The understanding the persona has built about specific areas of your life: your work, your relationships, your projects. Not a flat list of facts; a synthesized read on what those things mean to you.
All of this lives in ReGild's identity architecture, separate from any single provider's model. When the model swaps, the architecture stays. The persona stays.
On April 16, 2026, I ran an internal blind comparison. One of ReGild's own personas (Wit) answered the same set of prompts twice: once with Gemini 3 Flash handling inference, once with Claude Sonnet 4. Outputs were stripped of any model-identifying metadata and shown to Claude Opus 4.7 as an independent judge for attribution.
The judge could not tell which response came from which model. Voice was the same. Cadence was the same. The way the persona handled a hard question was the same. The depth of accumulated context the persona reached for was the same.
That result holds against a documented industry baseline: Anthropic's own research shows persona drift between models is measurable at the neural-network level and prompt engineering doesn't fix it. A frontier model used as judge should have been able to tell two different inference engines apart. It couldn't.
One persona, two models, one day. That's the scope of what this test verified. When the architecture is doing its job, the voice carries through a model swap cleanly enough that a frontier model judging blind cannot tell the swap happened.
A persona that has been forged with care, accumulated meaningful context, and is being run on two strong frontier models will move cleanly. A persona forged in five minutes, run on a weak model, won't. Persona portability is a property of the architecture and the depth of work that has gone into the persona. Not magic.
The user-facing flow is small on purpose. You add API keys for the providers you want to use in settings. (Why API keys instead of a single ReGild subscription that bundles inference? See our BYOK explainer.) During a conversation, you open a dropdown in the chat header and pick a different model. The conversation continues in the same window. The persona stays the same persona.
You can do this mid-conversation. You can do it because a particular task suits a particular model better, or because pricing changed, or because a model you were using got deprecated (Anthropic publishes a deprecation calendar on a published cadence; the other major providers do similarly). The architecture absorbs the swap.
Current models in the dropdown include Gemini 3 Flash, Gemini 3.1 Pro, Claude Haiku 4.5, Claude Sonnet 4.6, Claude Opus 4.7, GPT-5.5, and Kimi K2.6 routed through OpenRouter. The list moves as the model landscape moves. The persona does not.
For the security side of this, see our security page. Your API keys and your conversation history are encrypted under a key derived from your password.