Lstm output magic boundary

I use LSTM in RL model training
And in differents runs it has value range between 1 and 0 or between 0 and -1
Why does it have always 0 boundary ? Exploration process checks all values between -1 and 1 but LSTM output doesn’t, it always has 0 boundary
I use 1 layer lstm, with 10 input features, and 1 hid. state