The water balance of catchments will, in many cases, strongly depend on its state in the recent past (e.g. previous days). Processes causing significant hydrological memory may persist at longer timescales (e.g. annual). The presence of such memory could prolong drought and flood risks and affect water resources over long periods, but the global universality, strength, and origin of long memory in the water cycle remain largely unclear. Here, we quantify annual memory in the terrestrial water cycle globally using autocorrelation applied to annual time series of water balance components. These time series of streamflow, global gridded precipitation, and GLEAM potential and actual evaporation, along with a GRACE (Gravity Recovery and Climate Experiment)-informed global terrestrial water storage reconstruction, indicate that, at annual timescales, memory is typically absent in precipitation but strong in terrestrial water stores (root zone moisture and groundwater). Outgoing fluxes (streamflow and evaporation) positively scale with storage, and so they also tend to hold substantial annual memory. As storage mediates flow extremes, such memory often also occurs in annual extreme flows and is especially strong for low flows and in large catchments. Our model experiments show that storage–discharge relationships that are hysteretic and strongly non-linear are consistent with these observed memory behaviours, whereas non-hysteretic and linear drainage fails to reconstruct these signals. Thus, a multi-year slow dance of terrestrial water stores and their outgoing fluxes is common; it is not simply mirroring precipitation memory and appears to be caused by hysteretic storage and drainage mechanisms that are incorporable in hydrological models.@en