LSTM vector sum of inputs as memory

Question

I want to set up a long short-term memory (LSTM) network to have the vector sum of the inputs, which live in R^d, as its memory (ct). what is the required choices of the weight, and activation functions to do this?