Streamflow is typically autocorrelated, although the strength of this memory varies through time and between catchments. The degree to which prior conditions influence streamflow informs how we predict future streamflow and can drive temporal clustering of hydrological extremes. We introduce memory decay curves, which are exponential decay functions that summarize streamflow autocorrelation over time lags from one month to three years. These curves can be summarized using two parameters: initial memory and memory half-life. We fit these memory decay curves to weekly streamflow data from 9352 stations across Europe for the years 1980–2021 and identify four archetypes based on the combination of strong (or weak) initial memory and long (or short) memory half-life. These archetypes show distinct geographic patterns across Europe, with strong and long memory most common in regions with large aquifers or deep bedrock, including the Paris Basin and parts of Poland and eastern Germany. Weak, short-lived memory is widespread, especially in humid catchments. Streamflow memory, empirically and conceptually, relates to existing streamflow signatures – including the baseflow index and coefficient of variation – but explicitly represents memory across multiple timescales. This memory primarily originates through catchment processes rather than memory of meteorological forcing because the latter is generally weaker than and geographically unaligned with streamflow memory. Streamflow memory at monthly to yearly time steps consistently empirically links to climate characteristics, whereas memory more strongly connects to surface features at shorter timescales and to subsurface features at longer timescales. We compare memory of streamflow predictions from two hydrological models (mHM and EFAS) and find that modelled streamflow, on average, has weaker monthly memory but stronger yearly memory compared to observed streamflow. There also exist large regional variations in memory between models and between models and observations leading to both over- and under-estimations of memory. The streamflow memory decay curves presented here bridge commonly studied event and annual timescales in streamflow autocorrelation. Moreover, they identify regionally distinct memory behaviours across multiple timescales, which can help advance process understanding and hydrological modelling.