Continual learning (also called lifelong learning in the research literature) is what humans do naturally and current LLMs conspicuously do not: a deployed model’s weights are frozen, so it performs its thousandth code review with exactly the competence of its first. The technical obstacle is catastrophic forgetting, and the classical toolbox (weight penalties, replay buffers, per-task adapters) buys retention only by giving up some new-task learning.
The topic became a headline debate in 2025-26: Dwarkesh Patel and Andrej Karpathy argued that missing continual learning is the key bottleneck to broadly capable agents, while Dario Amodei countered that long-context in-context learning plus scale may be enough. Meanwhile production agents approximate it at the context level with self-written memory stores and scheduled consolidation, and research approaches like test-time training and sparse memory finetuning push learning back into the weights.