Hi Shubham,
The LSTM (Long Short-Term Memory) layer in a neural network is designed to remember values over arbitrary time intervals which indeed helps in maintaining and learning long-term dependencies. However, after training, the hidden and cell states of the LSTM layer are reset to zero. This is standard behavior for LSTMs, and it doesn't mean that the LSTM layer has not learned anything or that it's not working properly.
If you want to maintain the state of LSTM for some reason (like in case of time series prediction where you want the model to remember the state from the previous sequence), you can refer to the explanation for Open Loop Forecasting and Closed Loop Forecasting in the following documentation link:
Here "predictAndUpdateState" function has been used which updates the network state at every timestep.
Hope this helps!
