The mapping between neural activity and external behavioral variables can change across experimental conditions and behaviors. For example, in prefrontal cortex, the coding of a stimulus’s direction is different soon after stimulus presentation, versus during the delay period. Similarly, for a monkey performing a wrist movement task, the same linear mappings (decoders) cannot simultaneously predict movements when pushing against a spring, versus freely moving. While neural network-based decoders can have the capacity to accurately perform mappings across changing conditions, they are challenging to interpret. That is, we cannot simply look at a neural network that has been fit to decode from multiple experimental tasks, and understand whether, or how, the mapping has changed over tasks. Here, we developed a decoding architecture that enables flexible decoding while maintaining interpretability about how decoders change across time and behaviors. Our “Mixture of Decoders” model simultaneously learns multiple linear mappings (simple decoders), and at each time point, learns which of these (or which combination of these) mappings to use. This architecture lends itself to interpretability, as we can examine the cluster assignments across different time points and behaviors, and it also is built from simple linear mappings. We evaluated our Mixture of Decoders model across two new datasets, testing its ability to predict the relationship from motor cortex activity to muscle activity. In a naturalistic monkey cage dataset, we found separate clusters that specifically related to grasp and release submovements. In a freely-moving mouse dataset, we were able to discover how neural-behavioral mappings changed over wide-ranging behaviors, such as determining that eating/grooming use similar mappings, as opposed to walking. In both datasets, our approach outperformed decoding from manually labeled behaviors, and had performance comparable to flexible nonlinear methods, demonstrating the ability to achieve interpretability without sacrificing performance.