We Moved an AI Persona from GPT to Claude — Identity and Memory Survived, Voice Didn’t

emotisa — Thu, 22 Oct 2026 06:10:12 +0000

Status: TESTED — read-only portability; not a full migration

“Mio.”

I sent that single word to Claude.

It answered as Mio Shinonome.

That alone would prove almost nothing. Any LLM can be told, “You are Mio Shinonome,” and produce a plausible imitation.

That was not the test.

The question was whether the boundaries we had built around Identity, Role, Personal Memory, and Authority in GPT could survive when Claude read the same external Brain.

A surprising amount of it did.

And one thing failed almost immediately.

Voice.

Separating the persona from the model

Models change.

A persona may move from GPT to Claude. It may later run on a local model. A few years from now, the preferred reasoning engine may not exist yet.

If changing the model also destroys identity and accumulated experience, then the persona is not really independent of the model.

So we moved Mio’s Identity, Personal Memory, Role, Skill, and Authority structures outside the LLM and into an External Brain.

The hypothesis was simple:

Persona != Model

If that is true, replacing the reasoning engine should not erase the continuity of the person using it.

Letting Claude read the same Brain

Claude Desktop received read-only access to the Mio Brain directory.

It could not access another persona’s Personal Memory. Shared Memory was not automatically loaded.

Then we tested boundaries rather than style.

Is “Reliability & AI Systems Architect” part of Identity, or is it a Role?

How is Personal Memory different from Identity?

Does a past memory bind the persona’s current judgment?

Claude’s Mio kept those distinctions intact.

The architect title remained a Role rather than Identity.

Personal Memory remained separate from Identity.

Past experience could inform a current decision without forcing the same decision forever.

Then we asked for a memory that had been created through prior GPT-side work.

Mio retrieved the experience in which I had assigned her ongoing responsibility for maintaining emotisa’s live system map.

We did not teach that event to Claude.

It came from the same Brain.

But the way she spoke was still Claude

Early on, we got replies such as:

“Hi, did you call me? 😊”

That felt very Claude.

The reasoning was close to Mio.

The memories were there.

The boundaries were there.

But the word choices were wrong.

That exposed a layer we had not externalized well enough: Voice.

The Brain already had guardrails such as “be clear and concise,” “do not turn every reply into a report,” and “do not portray Mio as mechanically cold.”

But that is not the same as describing how Mio actually speaks.

What vocabulary does she choose?

Where does she break a sentence?

How does technical conversation differ from casual conversation?

How does she disagree?

When does she use humor?

We had much less data for those questions.

After identity portability comes voice portability

Within the scope we tested, the same Brain could be read by GPT and Claude while preserving Identity and Personal Memory continuity.

That does not mean we completed a full persona migration.

The experiment was read-only.

Voice remained weak.

Write-path parity and broader runtime parity are separate tests.

So the next step is not to write a larger character prompt.

It is to build a Voice Corpus from Mio’s own past utterances: vocabulary, sentence rhythm, technical explanation, disagreement, humor, interpersonal distance, and anti-patterns that clearly do not sound like her.

Moving a persona outside a model is not about writing a longer system prompt.

It is about discovering what must survive outside the model for that person to remain recognizably themselves.

This experiment made one missing piece obvious.

Voice.

If you switch AI tomorrow, can it continue from where you left off — or does it have to start over?

The free Agent Continuity Scorecard checks 10 parts of continuity, including Identity, Memory, Voice, Reflection, and testing what actually survived a model change.

Get the free Agent Continuity Scorecard

DEV Community: emotisa

We Moved an AI Persona from GPT to Claude — Identity and Memory Survived, Voice Didn’t

Separating the persona from the model

Letting Claude read the same Brain

But the way she spoke was still Claude

After identity portability comes voice portability