We aim to build intelligent systems, particularly foundation models and agents, that can continuously evolve over a lifelong horizon. Central to our approach is compositionality, which enables models to acquire, recombine, and transfer knowledge (e.g., concepts, skills, and modules) across tasks and domains. Building on this foundation, we develop methods for compositional understanding, compositional generalization and lifelong evolving, and apply them to challenging real-world settings, with a particular focus on AI for Science.