Online continual learning (OCL) systems must learn from a stream of tasks without revisiting past data, and they often lose previously acquired knowledge in the process. This phenomenon, known as catastrophic forgetting, has motivated many heuristic fixes, but the underlying problem has lacked a unifying formal treatment.

The new preprint, posted on arXiv, takes a step toward filling that gap by proposing a theory of forgetting specifically for the OCL setting. Rather than presenting another method, the authors aim to formalize what forgetting is and how it can be measured, potentially giving researchers a common language for comparing approaches.

Because the abstract is brief, the full scope of the theory is not yet clear from the announcement alone. Still, the emphasis on formalization suggests a move toward more principled analysis in a field that has largely been driven by empirical trial and error. If the framework holds up, it could sharpen both the diagnosis of forgetting and the design of more robust continual learners.