Anticipating the next item a user will engage with is vital for the success of recommendation systems, particularly given the ever-changing nature of user choices. In this research, we introduce a session-based attention recommendation model (SARM) that leverages Gated Recurrent Units (GRU) enhanced with an attention mechanism to adeptly capture complex temporal dynamics within a session, allowing for precise and efficient next-item predictions. Initially, the embedding layer maps one-hot encoded item features into dense, continuous vectors. The sequence of these embeddings is handled by the sequential modeling layer using GRUs to identify temporal patterns through reset and update gates. Each hidden state is given a dynamic weight through the attention mechanism, which creates a context vector that aggregates the most pertinent information from the sequence. The prediction layer then uses this context vector to compute next-item predictions by a linear transformation followed by a softmax function. The results of experiments demonstrate that our approach outperforms the other baselines in terms of improving the accuracy and ranking quality of next-item prediction.