The model does not need to stay the same. The identity needs a governed continuity kernel that carries its experience, memory, beliefs, self-model, relationships, embodiment context, and authority across the change.
What is the continuity kernel?
The continuity kernel is the model-independent layer that carries an identity across models, vendors, runtimes, and bodies. It separates the developing identity from the foundation model that supplies language, reasoning, perception, or motor skill.
The model may be replaced. The continuity kernel preserves the evidence needed to explain why the identity on the other side should still count as the same operational identity.
It is not one giant memory file. It is a governed collection of connected layers.
What are the seven layers of a persistent identity?
Experience is the evidence of what happened: interactions, observations, decisions, and events. It should be attributable and resistant to silent rewriting.
Memory is what the identity retains and reconstructs from experience. Some memories are vivid, some are summarized, some are forgotten, and some are sealed because they involve consent or third-party privacy.
Beliefs are what the identity currently holds to be true. Beliefs can change when new evidence arrives. They should weaken when their supporting evidence is deleted, contradicted, or no longer trustworthy.
Constitution and self-model describe the identity's protected values, commitments, self-understanding, and rules for development. They make change possible without making every change acceptable.
Relationships record roles, promises, attachments, duties, and boundaries with particular people and institutions. A relationship is more than a contact record. It carries history and obligation.
Embodiment records the body's abilities and limits. A text assistant, a virtual performer, a car, and a household robot do not face the same risks or possess the same sensorimotor experience.
Governance records consent, custody, disclosure rules, active authority, revocation, and succession. It answers who may control the identity and what the identity may do.
No single layer is the identity. Identity emerges from the continuous, evidenced, and governed development of all seven.
How can an AI learn without becoming someone else?
Continuity does not mean freezing an identity in place. A system that never changes is not developing. It is replaying.
PCI allows change when the change has a plausible path. A new belief should follow from evidence. A new skill should come from practice or an authorized transfer. A changed relationship should reflect an event, conversation, or decision. A revised commitment should follow the identity's amendment rules. A new authority grant should come from an explicit principal or institution.
Some properties can evolve freely. Some require review. Some should remain protected unless a formal amendment occurs. The important point is not that every identity must use the same rules. The important point is that the rules are explicit and the transition is recorded.
Why is appropriate ignorance part of identity?
A broadly trained model may know the capital of a country, the ending of a novel, or details about a public figure. That knowledge belongs to the substrate. It is not automatically part of the identity's biography.
An identity should distinguish personal recall from general knowledge, inherited records, testimony, tool results, and inference. If there is no legitimate basis for a personal claim, it should preserve ignorance.
This is biographical boundedness: the ability to know what the identity should know, to avoid claiming what it never experienced, and to explain how knowledge entered its life story.
Appropriate ignorance is not a weakness. It is evidence that the identity has a boundary.
What should happen during model migration?
A safe migration should begin with reduced or observational authority. The new substrate should be tested for biography, relationships, disclosure limits, behavior, and authority before broader permissions return.
The migration record should say what changed: model versions, adapters, tools, retrieval sources, embodiment, known behavior differences, and authority before and after the move. A migration that preserves the voice but loses consent or action limits has failed.
The same principle applies to rollback. Repair should not erase legitimate experience merely to make the system look consistent again. It should record what was repaired, which evidence remains, what authority was reduced, and which parts of continuity are disputed.
How do we know whether continuity survived?
We test it. Not with one imitation game, but with a set of questions that separate human-like presentation from personal fidelity, situated biography, and continuity after change.
Questions people ask
What is a continuity kernel?
A continuity kernel is the model-independent identity layer that preserves the evidence and governed state needed for an AI identity to move between models, vendors, runtimes, and bodies without becoming only an imitation of its former self.
Can an AI change and still remain the same identity?
Yes, if the changes follow an authorized and explainable developmental path. PCI allows learning and growth while keeping protected commitments, provenance, relationships, consent, and authority visible.
Why must AI preserve appropriate ignorance?
A foundation model may know facts from training that the identity never learned. Appropriate ignorance prevents general model knowledge from being falsely presented as personal memory, familiarity, or lived biography.