It is essential to identify and classify human emotions via the deep learning with computers. Therefore, Electroencephalogram (EEG) is used as source of emotions of human body. In this paper, a long short-term memory (LSTM) model is proposed for classification of positive, neutral, and negative emotions. This model is applied to a dataset involved in 3 classes of emotions with a total 2100 EEG samples from two persons. The proposed model is trained using TensorFlow library with a tuning method to achieve highest accuracy of emotions prediction. Receiver operating characteristic (ROC) curve is used to compute the model performance. The results demonstrate that the LSTM provides high performance in the classification of human emotions. Furthermore, The LSTM model has a testing accuracy of 98.13%, a macro average precision of 98.14%, and the area under the ROC curve for this model is 100%.