In dynamic positron emission tomography (PET) reconstruction, the importance of leveraging the temporal dependence of the data has been well appreciated. Current deep-learning solutions can be categorized in two groups in the way the temporal dynamics is modeled: data-driven approaches use spatiotemporal neural networks to learn the temporal dynamics of tracer kinetics from data, which relies heavily on data supervision; physics-based approaches leverage \textit{a priori} tracer kinetic models to focus on inferring their parameters, which relies heavily on the accuracy of the prior kinetic model. In this paper, we marry the strengths of these two approaches in a hybrid kinetics embedding (HyKE-Net) framework for dynamic PET reconstruction. We first introduce a novel \textit{hybrid} model of tracer kinetics consisting of a physics-based function augmented by a neural component to account for its gap to data-generating tracer kinetics, both identifiable from data. We then embed this hybrid model at the latent space of an encoding-decoding framework to enable both supervised and unsupervised identification of the hybrid kinetics and thereby dynamic PET reconstruction. Through both phantom and real-data experiments, we demonstrate the benefits of HyKE-Net -- especially in unsupervised reconstructions -- over existing physics-based and data-driven baselines as well as its ablated formulations where the embedded tracer kinetics are purely physics-based, purely neural, or hybrid but with a non-adaptable neural component.