How to clear TF/Keras memory when loading a new model while other models are still in memory and in use?

Question

I am writing a server application that allows TCP clients to ask for predictions based on a previously loaed TF/Keras model. The TCP client is allowed to load such a model b