A better model arrives. A new application fits the work. You decide to switch.
The intelligence improves, but the context you built may be left behind: project organization, decisions, preferences, corrections, and the reason something changed.
If you have spent months explaining your world to an assistant, that is a meaningful loss. The important asset is the understanding that accumulated while you used the tool.
An export needs its meaning
In examining multi-year conversation exports, this project found a specific continuity gap: the conversations came back without the links that connected them to their projects. One export included project definitions separately.
The words were there. The organization was harder to reconstruct.
That observation shaped the work. A usable exit needs more than a pile of text. It needs the information that lets a new system understand what the records refer to, which decision was accepted, what changed, and what remains uncertain.
The finding describes the exports examined. It is not a claim about every provider or every future export format.
Separate the record from the view
A summary can be useful. So can an embedding, an index, or a graph.
Each is a way to work with the evidence. The source material and its meaning need a durable home as well. Otherwise a transformation can become the only surviving version of what happened.
Sovereign Memory Protocol is being developed around provenance, custody, authority, integrity, and portability. Memory Core implements and tests mechanisms in PostgreSQL. Applications and connectors have their own responsibilities.
This separation makes replacement a design goal. A model or retrieval engine can improve without automatically becoming the owner of every record it touches.
Corrections should remain understandable
Suppose a shared project record says that an appointment was accepted. Later, the person explains that it was only proposed.
The new answer should reflect that correction. The history should still explain how the earlier belief arose and why it changed. A later assistant should be able to distinguish the proposal from the accepted commitment.
The same pattern applies to a household preference, a business decision, or a maintenance task. What was observed, inferred, proposed, and accepted can lead to different actions.
A confident model response cannot supply those distinctions by itself.
Ownership needs a practical exit
Our intended result is knowledge that remains understandable and useful across changes in tools and providers.
Core's alpha package-and-restore rehearsal is a bounded step toward that goal. It uses representative synthetic data. Live recovery and independent acceptance remain open work.
That limit belongs beside the result because portability has to be demonstrated for the relevant system. Owning an account or hosting a database does not finish the job.
The direction is durable: let models compete to be useful, while the person retains the evidence, decisions, corrections, and history that make the help personal.