Hydrological drought (HD), marked by prolonged low streamflow and depleted water storage, presents major challenges for water resource management. We introduce an interpretable deep learning framework that combines an attention-augmented Long Short-Term Memory network with Expected Gradients analysis to identify key drivers of HD across the contiguous United States (CONUS). Leveraging an 83-year (1940-2022) ERA5 data set, the model incorporates monthly precipitation, potential evapotranspiration, temperature, surface pressure, soil moisture, snow depth, and the Ni & ntilde;o-3.4 index to predict HD intensity in 547 minimally disturbed catchments from the GAGES-II database. Our analysis reveals three distinct HD mechanisms: (a) rapid-onset droughts driven by short-term hydroclimatic memory, (b) gradually evolving droughts shaped by cumulative hydroclimatic stress, and (c) climatically reinforced droughts, predominantly in arid regions, where large-scale teleconnections such as La Ni & ntilde;a amplify moisture deficits. Spatial patterns indicate that short-term memory-driven HDs dominate humid eastern US catchments, gradually evolving HDs are prevalent in the central Great Plains and Rocky Mountains, and climatically reinforced HDs are most frequent in the arid Southwest. Arid catchments are strongly affected by ENSO-driven precipitation deficits; semi-arid zones experience multiple drivers, including temperature anomalies and persistent precipitation deficits; subhumid catchments are primarily controlled by precipitation deficits and SM depletion; and humid regions face abrupt drought onset from sharp precipitation declines, often compounded by snowmelt timing in snow-dominated areas. This framework offers a robust foundation for seasonal drought forecasting and adaptive water management strategies.