We revisit the architecture of existing temporal graph learning methods and propose a simple yet effective framework without memory or aggregation for future link prediction.