Large language models have shown promise in assisting humans, but most real-world AI remains reactive, responding only to explicit user prompts. Current proactive AI agents often fail in multimodal reasoning over long-term memory, which requires agents performing temporal reasoning over long-horizon multimodal histories and continual knowledge update, and thus struggle to predict users’ needs and intent. To address this challenge, we present ProactiveEngine, the first foundation model designed for proactive multimodal agents, which generates proactive responses in long-horizon settings. ProactiveEngine supports both video plus text and text-only input. Our model achieves an F1-score of 83.3%, substantially outperforming current commercial models. Our model also achieves strong performance on the public Proactive Benchmarks in a zero-shot setting without further training, while enabling a 7.8× increase in token context window and over 100× more historical events for long-horizon proactive prediction. We also introduce NeuroAuroraBench, a multimodal benchmark designed to evaluate long-term memory proactive prediction. The benchmark resulted in 24.8k annotated sessions capturing dynamic changes.