Abstract: This article reviews the evolution of continual learning from the study of catastrophic forgetting to autonomous evolution in open environments. It summarizes the main approaches developed for stability preservation and discusses the new changes brought by large models, including parameter-efficient fine-tuning, modular composition, external memory, tool use, and learnware reuse. On this basis, the article introduces the engineering perspective of continuous development, continuous integration, and continuous deployment, and argues that continual learning is extending from a parameter update problem of a single model to a system-level problem of capability development, integration, and online deployment in open environments.