Blood pressure (BP) is a crucial indicator of cardiovascular health. This paper introduces a new approach for non-invasive BP (NIBP) measurement using photoplethysmogram (PPG) signals. The proposed method utilizes a memory-based stacked autoencoder (MemSAR) to extract features from PPG time-series data, effectively capturing temporal dependencies between consecutive cardiac cycles. These extracted features are then applied to various regression models. It was found that MemSAR combined with k-nearest neighbors (kNN) model provides balanced accuracy with a mean absolute error (MAE) of 2.60 mmHg for systolic BP and 1.95 mmHg for diastolic BP, among other popular regressors, validated on the MIMIC-III Waveform Database. This approach also provides better results over few recent published works using machine learning and deep learning for NIBP measurement.