Different service providers use various recommender systems to suggest items to users. Recently, research has focused on using Recurrent Neural Networks (RNNs) for recommendation systems due to their ability to utilize past session data for current recommendations. This thesis proposes a new architecture that combines two RNNs: one storing past session data and another storing current user interactions, to enhance item recommendations in the current session. The model is pre-trained using item-interaction learning, trained on a GPU, and employs the Dropout technique to reduce overfitting. Parameters such as batch size, epoch size, and dropout layers were considered in designing the RNN. The results were compared and analyzed to identify the best architecture for accurate recommendations based on past sessions. The approach involves using a second RNN layer to learn from past sessions and predict current user interests, with the first RNN storing current sessions. Item-interaction embedding was used to determine the likelihood of co-occurring item interaction pairs. The models performance was tested on three different datasets, showing improved recommendations based on past session interactions.