A study comparing Transformer and LSTM models for predicting streamflow in ungauged basins found that LSTM outperformed the Transformer, particularly when incorporating downstream information. This suggests that recurrent memory is more effective for hydrologic sequence inference than encoder-only architectures.