In recent years, long short-term memory (LSTM) networks equipped with static catchment attributes have become the standard tool for prediction in gauged and ungauged basins. However, the real effectiveness of static catchment descriptors in guiding prediction in unseen basins has sparked considerable debate. In particular, it remains unclear whether catchment attributes truly allow knowledge transfer to sites where the model has not been trained. Furthermore, it is not clear how much information attributes really contain for streamflow prediction and how much they still miss, whereas recent studies show that this information is mainly related to climate. We address these questions by training an encoder-based LSTM conditioned on the weather. This way, the architecture encourages the encoder to prioritize attribute information complementary to the meteorological forcing variables. Remarkably, our findings reveal that the predictive information of commonly used attributes can be represented by two latent features in most of the catchments. Further, we relate these latent features to the topography and the vegetation of the catchment, bridging the gap between performance and interpretability.