Primitives of Memory: Foundations of Continual Learning Algorithms

Tea Talks July 31, 2026 After a decade of rapid growth, the continual learning literature is vast: replay, regularization, gradient projections, low-rank adapters, modular architectures. Even the newest LLM training pipelines rediscover each of these mechanisms under new names. In this talk I present a theoretical account of this landscape. I will organize the algorithms into a map: one general optimization template, two dual method families, mirrored failure modes, and a shared computational limit. I will then show that beneath their variety, most of these algorithms implement a single princi