It’s technical limitations. I haven’t found out a method to optimize the model for short term memory.