Human mobility forecasting is pivotal for urban applications like transportation planning and personalized recommendations. Yet, existing approaches face significant challenges in generalizing to unseen users or locations and effectively capturing dynamic user intentions, primarily due to data scarcity and the inherent complexity of spatio-temporal mobility patterns. To this end, we propose a framework, Zero-Shot Human Mobility Forecasting via Large Language Model with Hierarchical Reflection (ZHMF), combining a semantic-enhanced retrieval-reflection mechanism with a hierarchical language modeling system. The task is reformulated as a natural language question-answering paradigm. By leveraging LLMs’ semantic understanding of user histories and context, our approach can effectively handle unseen prediction scenarios. A key innovation is our hierarchical reflection mechanism, which decomposes the forecasting process into an activity-level planner and a location-level selector. This design facilitates collaborative modeling of long-term user intentions and short-term contextual preferences through iterative reasoning and refinement. Extensive experiments on real-world datasets demonstrate that ZHMF significantly outperforms state-of-the-art models. Ablation and case studies further validate the efficacy of our design in capturing nuanced user intentions and adapting to diverse contexts.