Can we trust ocean memory trends?

Ocean memory, defined as the persistence of sea surface temperature (SST) anomalies due to the ocean’s high heat capacity, can last from a few days to several years, affecting both atmospheric and oceanic conditions. Previous research suggests a substantial historical increase in ocean memory. Here we reproduce and reassess this ocean memory increase using multiple satellite-based and reanalysis SST datasets. While all datasets agree on an increase in ocean memory, the magnitude of the trend and its spatial pattern vary considerably. To further interpret these differences, ocean memory is decomposed into high (1–15 days), intermediate (15–365 days), and low (longer than one year) frequency components. This decomposition reveals that the trend from the intermediate-timescale component dominates the historical increase in total ocean memory across all datasets. 

Cross-dataset discrepancies in the magnitude and spatial structure of trends in total ocean memory most likely reflect uncertainties in SST estimation, including structural and parametric differences in product construction, rather than genuine physical variability, and may include non-physical artifacts such as inhomogeneities or processing biases. Here we further compare all estimates derived from SST products with that from mooring records, which provide an independent observational benchmark. However, the relatively short temporal coverage of the mooring data makes the comparison inconclusive. In addition, we also evaluate ocean memory trends using a free-running ocean model forced with atmospheric reanalysis (ERA5 and JRA55) which shows positive trends, although area-averaged trend magnitude are 4 to 11 times weaker depending on which SST dataset is used for comparison. In contrast, CMIP6 models show no consistent evidence of such a positive trend, even for future periods under high-emission future scenarios (e.g., SSP5-8.5).   

Some satellite-based and reanalysis SST products also contain data artifacts that may influence these trends. For example, dramatic jumps are obvious in temporal autocorrelation in several datasets, which are clearly unrelated to any physical process or climate driver. Overall, the results point to an increase in ocean memory trends yet highlight uncertainties arising from dataset artifacts, limited mooring benchmarks, and the absence of robust model support.