This thesis addresses the challenge of efficiently processing ultra-high-speed (microsecond-level temporal resolution) event-based camera data. Due to their asynchronous, sparse, and spatially unordered nature, traditional frame-based image processing methods are ineffective for handling such data streams. To overcome these limitations, we propose Eventformer, a novel event-based representation learning framework. Eventformer employs a self-attention mechanism to effectively model spatial interactions among events and incorporates an associative memory-augmented recurrent module to dynamically store and retrieve latent states specifically at event locations. This approach substantially reduces computational costs while preserving high performance on standard event-based object recognition benchmarks. Expanding upon the Eventformer framework, we develop a more comprehensive spatiotemporal representation learning framework tailored specifically for event-based dense prediction tasks, which require precise pixel-level associations. Our method uniquely manages sparse and asynchronous event streams by projecting them into hierarchical, multi-level latent memory structures through an attention-based spatial association. Additionally, a data-adaptive memory update mechanism selectively refreshes memory structures only when significant new information is identified, significantly reducing latency and computational overhead while maintaining competitive performance. Furthermore, we demonstrate the practicality and robustness of our proposed framework in complex, real-world scenarios such as high-speed autonomous navigation and obstacle avoidance, integrating compute-adaptive perception and control. In summary, this research introduces a novel memory-augmented representation learning pipeline that significantly enhances the efficiency and effectiveness of real-time processing of event-based camera data, supporting a variety of perception and control applications with notable improvements in computational efficiency and reduced latency.