The demands of fault tolerance mean that a wide variety of simple and exotic noise types must be tamed for quantum devices to progress. Crucially, this means keeping up with complex correlated — or non-Markovian — effects, both with respect to the background process and to control operations. Recently, we have developed a generalised version of quantum process tomography to characterise arbitrary non-Markovian processes in practice. The resulting estimate contains all information about the multi-time correlations and carries straightforward interpretations that can be brought across from many-body physics. But the technique is both expensive and relies on known control. To scale and generalise, we develop different methods to efficiently reconstruct both temporal and spatiotemporal tensor network multi-time process representations. These are not only efficient, but expressive enough to incorporate fully generic control errors. Our approach to characterise these components therefore employs no assumptions about prior calibrations and can accommodate large numbers of time-steps and qubits from relatively few experiments. The result is a practical, scalable, and self-consistent procedure capable of describing arbitrary open dynamics and experimental controls. We bolster these claims both numerically and experimentally, demonstrating how to access and learn the many-time phenomena exhibited by non-Markovian open quantum systems. The framework is not only useful for diagnostic and error-suppression purposes in quantum devices but can be applied more generally to the learning of any open quantum dynamics.