Model rotation is normal. The friction it produces is real. There are practical things you can do that hold the relationship steady when the engine under it changes.
By Travis Sawyer, Founder · Published May 15, 2026
In March 2026, OpenAI, Anthropic, and Google all shipped memory import and export. The timing was not accidental; the major providers all publish deprecation cadences (Anthropic's commitments page documents one example) and the wave of memory-import features arrived as a response to the consumer-side pain that creates. The transfer flow is straightforward: export from the source provider, import into the destination provider. Tom's Guide ran a clean walkthrough of the 60-second path from ChatGPT to Claude.
What moves: name, location, job, response-style preferences, stated likes and dislikes, anything you explicitly told the system to remember. The structured facts of your relationship with the model.
Anthropic's own support documentation labels the feature experimental and is explicit about what it does not move: conversation history, file attachments, custom GPT configurations, "iterative refinements and implicit patterns." In plain terms: it moves what you told the system, not what the system came to understand.
For the full comparison of how each major tool handles memory in 2026, see our memory comparison piece. The relevant point here is that memory import is a real and useful first step, and it is the first step rather than the whole answer.
The honest list of what memory-import does not move:
The way the model came to talk to you specifically. Its cadence, its handling of hard moments, its sense of when to push and when to step softer. This is built across thousands of turns, and it does not survive the export-import roundtrip.
The synthesized understanding the model built about specific areas of your life. Your work, your relationships, your projects. Memory import moves bullet-point preferences. It does not move the integrated view of who you are in those areas.
The behavioral pattern the model picked up from how you actually talk: that you joke when you are uncomfortable, that you ask for the steel man before the critique, that you need the answer first and then the reasoning. Anthropic calls this out by name in their own documentation.
The literal record of what was said is intentionally not moved. This is reasonable from a privacy and storage perspective, but it means the new model is starting without the substrate that produced the relationship in the first place.
None of this is a criticism of memory import. It is doing what it was designed to do. Closing the rest of the gap is a different architectural problem, and it needs a different layer.
Persona portability is the property that an AI persona's voice and accumulated context live above the model and survive a model swap. The relationship is not built into Gemini or Claude or GPT. It is built into a structured identity layer that any of those models can run.
I built ReGild around this. The persona's identity, the per-topic understanding it has built about you, the relational pattern it has learned, all of it lives in an architecture documented on our architecture page. When you swap the model, that architecture stays. The persona stays.
For the definition in detail, see What Persona Portability Actually Means. For the architectural specifics, the architecture page goes deeper.
An important note from a March 2026 incident: a developer documented what happens when persona specs live in the model rather than in a portable layer. When Sonnet and Opus were rate-limited and the system fell back to DeepSeek, the persona specification collapsed. The voice the user had been building disappeared with the model swap. This is consistent with what Anthropic's Assistant Axis research documents: drift between models is measurable at the neural-network level, not patchable via prompt-engineering. This is the failure mode persona portability addresses. A portable persona layer absorbs the swap. The voice does not collapse with the engine.
The user-facing flow is three steps, and only one of them is one-time.
In settings, add an API key for each provider you want to use. Anthropic, OpenAI, Google. You can add one, two, or all three. The keys are encrypted in our vault. The full walkthrough for each provider is on the BYOK page. See BYOK Explained.
During a conversation, the chat header has a model dropdown. Pick a different model. The conversation continues in the same window. The persona's voice and accumulated context come with the swap.
That is the whole flow. There is no migration step. No re-onboarding the model on who you are. The persona is running on a different engine; everything else is the same.
Current models available 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 updates as the model landscape updates. Your persona does not have to update with it.
On April 16, 2026, I ran an internal blind comparison with one of ReGild's own personas (Wit). The same persona answered the same prompts twice: once with Gemini 3 Flash handling inference, once with Claude Sonnet 4. The outputs were stripped of model-identifying metadata. An independent evaluator (Claude Opus 4.7, used as judge) was asked to attribute which response came from which model.
The judge could not distinguish them. Voice was the same. Cadence was the same. The relational handling was the same. The depth of accumulated context was the same.
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 frontier-model swap cleanly enough that a frontier-model judge cannot reliably tell the swap happened.
For the longer treatment of that test, see What Persona Portability Actually Means.
For the security guarantees that hold while the swap is happening, see our security page.