A key metric for improving business cafeteria operations, inventory control, and customer satisfaction is accurate sales forecasting. k-Nearest Neighbors (k-NN) and other standard machine learning algorithms usually fail to identify temporal patterns in sales data, which leads to predictions that are not accurate. We provide a better forecasting technique that overcomes this limitation by utilizing Long Short-Term Memory (LSTM) networks, a specific kind of recurrent neural network that can recognize long-term dependencies in sequential data. The performances of LSTM and k-NN were compared and evaluated using a Kaggle corporate cafeteria sales dataset. Based on experimental results, the LSTM model outperformed k-NN with a mean prediction accuracy of 87.57% compared to 83.99%. The significance of the improvement (p = 0.038) was validated by statistical testing with an independent samples t-test.