There has been recent success in learning from static graphs, but despite their prevalence, learning from time-evolving graphs remains challenging. We design new, more stringent evaluation procedures for link prediction specific to dynamic graphs, which reflect real-world considerations and can better compare different methods' strengths and weaknesses. In particular, we create two visualization techniques to understand the recurring patterns of edges over time. They show that many edges reoccur at later time steps. Therefore, we propose a pure memorization baseline called EdgeBank. It achieves surprisingly strong performance across multiple settings, partly due to the easy negative edges used in the current evaluation setting. Hence, we introduce two more challenging negative sampling strategies that improve robustness and can better match real-world applications. Lastly, we introduce five new dynamic graph datasets from a diverse set of domains missing from current benchmarks, providing new challenges and opportunities for future research.