MULTIPLE datasets (input-target) to train a SINGLE Neural Netwok

Hi! I'm trying to build a NARXNET to make time series prediction. I have different input-target pairs available to make training.

I read that is not possible to "retrain" a network with a new input-target pair because at the beginning of each training, initial condition are randonmly re-written , so there is an -overwrite- and not an -update- of the network. Is it right??

So, Is there a way to use different training data pairs on the same network?

I tried to brutally concatenate different pairs -->newinput=[input1 ; input2] // newtarget=[target1;target2], but but my fear is that the discontinuity between the signals can cause network problems.

N.B I Have another problem during training using train_function like "trainbr" and "trainlm" the training stop very early, It reaches low value of best_validation_perfomance at 10-15 epoch than the three curves of training validation and test abruptly diverge leading no more training improvement. Any suggestion??