What can survive a change of mind?
Storage is cheap. A usable representation is not. How can learning remain inspectable and revisable when the model, the tools, or the people around it are replaced?
Research
Questions, not findings.
This page describes an area of inquiry. It is not a paper list, a lab report, or a claim that the work is done. Neighboring fields already name parts of it: continual and lifelong learning, distillation and adapters, retrieval and external memory, replay, model merging. We do not claim those fields. For the long form, read Intelligence should accumulate.
Current questions
How does later work begin from earlier work?
Inheritance that cannot be examined is a rumor. What interfaces would make prior distinctions the default starting point for a new agent, a new model, or a new person, without pretending the past is always right?
When is keeping it actually cheaper?
An archive is not a tradition. Under what conditions does later work get clearer, cheaper, or more ambitious because earlier work was kept — and how would we notice if it did not?
What should remain local?
Privacy, a failed path that should die, a distinction that is only true in one room. Locality is not the same problem as revision. We leave the boundary open.