TLDR: A model can use new knowledge in context and still lose it the moment the session ends, so persistent learning depends on deciding how experience should change the model, and who makes that decision. SEAL hands part of that job to the model itself: it generates its own self-edits, applies them, and is trained with reinforcement learning on how much the updated model improves, lifting question-answering accuracy in one knowledge setting from 32.7% to 47.0%, versus 33.5% from training on the raw passage.
We separate remembering from learning across context, external memory, and parameter updates, look at why a training example is really a choice about computation, and at what LoRA does and doesn’t solve. We then turn to Nested Learning and Titans, which spread adaptation across different timescales, the catastrophic forgetting problem every persistent learner has to manage, and what evidence would show a system getting better at becoming better.
Imagine teaching a coding assistant a new API. During the conversation, it uses the documentation correctly. Open a fresh session without that documentation, and the lesson may disappear. The assistant displayed an ability in context, but the underlying model did not necessarily acquire it. For a system expected to improve through experience, that difference is fundamental.
The next question is who decides how an experience should change the model. Researchers usually design that process themselves. Work such as SEAL gives the model a role in constructing its own updates, while Nested Learning examines how adaptation can operate across different timescales. Together, they provide useful ways to think about persistent learning and the controls that make it beneficial.
Remembering and learning are different operations
Read more