The initial stages of the evolution of an open quantum system encode the key information of its underlying dynamical correlations, which in turn can predict the trajectory at later stages. We propose a general approach based on non-Markovian dynamical maps to extract this information from the initial trajectories and compress it into non-Markovian transfer tensors. Assuming timetranslational invariance, the tensors can be used to accurately and efficiently propagate the state of the system to arbitrarily long time scales. The non-Markovian transfer tensor method (TTM) demonstrates the coherent-to-incoherent transition as a function of the strength of quantum dissipation and predicts the non-canonical equilibrium distribution due to the system-bath entanglement. TTM is equivalent to solving the Nakajima-Zwanzig equation, and therefore can be used to reconstruct the dynamical operators (the system Hamiltonian and memory kernel) from quantum trajectories obtained in simulations or experiments. The concept underlying the approach can be generalized to physical observables with the goal of learning and manipulating the trajectories of an open quantum system.Introduction.-The dynamics of large open quantum systems are of interest to a broad range of disciplines, including condensed matter physics, ultrafast spectroscopy, and quantum information technology, just to name a few. Of particular interest is the interaction between the system under study and the environment to which it couples. Within the fast bath approximation, the evolution of the system's density matrix is dictated by a Lindbladian superoperator and can be regarded as a linear Markovian process. Nevertheless, in general the quantum trajectory of the open system is entangled with the bath and is therefore temporally correlated, i.e., non-Markovian. The analysis and simulation of this correlation is a daunting task, which often requires resources that scale exponentially with the system size. The root of the problem is the lack of a compact but complete representation of the information encoded in open quantum trajectories. The standard approaches fall into two classes: quantum master equations and path integral simulations. The first class of approaches is based on formally exact equations of motion, such as the Nakajima-Zwanzig formalism [1][2][3] or others, but restricted to either weak damping, high-temperature, short memory time or short simulation time [4][5][6][7][8][9][10]. The second class of approaches adopts the harmonic bath assumption which renders the use of stochastic Gaussian sampling or influence functional possible [11,12], but does not converge well with the system size, the length of the memory time or the strength of the dissipation. To overcome these difficulties, we need a radically different approach to dissipative quantum dynamics.