This paper introduces Cache-Augmented Generation (CAG) as an alternative to Retrieval-Augmented Generation (RAG) for knowledge-intensive tasks using long-context LLMs.