Autonomous multi-agent swarms play a vital role in post-disaster emergency search and rescue missions; however, in severely occluded unstructured environments like ruins, they remain restricted by the dual constraints of limited individual field-of-view and inter-agent collision risks. Traditional deep reinforcement learning methods often struggle to balance exploration efficiency with safety, leading to catastrophic collisions or excessive avoidance. To address these issues, this paper proposes a Hierarchical Safety Cooperative Control Framework integrating spatiotemporal memory and optimization theory. This framework decouples multi-agent coordination into two levels: a memory-based high-level policy utilizing gated recurrent units and centralized training with distributed execution to extract key spatiotemporal features from historical observations, mitigating perceptual uncertainty; and an optimization-based low-level safety layer employing control barrier functions. The latter constructs a real-time safety filter formulated as a quadratic programming problem to constrain agent states within a safe invariant set without retraining, ensuring theoretical safety guarantees. Simulation results across scenarios with 10 to 20 obstacles demonstrate robust performance: task success rates range from 87.6% to 93.4%, representing an approximate 17.8% improvement over baselines, while collision rates decrease by 41.1% to 52.1%, and average rewards increase by 29.1% to 33.1%. These findings validate the framework's scalability, robustness, and adaptability in heterogeneous, dense obstacle environments, offering a reliable solution for safe and efficient autonomous cooperative rescue.