What is the best way to save Keras RNN layer states during training and prediction?

Question

I have an LSTM processing an infinite stream of events coming from multiple users. Thus, events are processed one timestep at a time as they arrive. Each user might behave diffe