In this paper, we extract a range of features including time-based, spectral-based, and phase-based to characterize working memory load in EEG recordings from a reading task in which different levels of working memory load were induced. It is demonstrated that a subset of time-based and spectral-based features - the mean, cross-correlation, and energy of the EEG signals - recorded from a few frontal channels in the delta frequency band, and also the statistics of selected wavelet coefficients are representative of working memory load and change most consistently in accordance with the induced load. We show classification accuracy of up to 100% for three working memory load levels across all five subjects. This is achieved using a multi-class support vector machine (SVM) trained on the above features from four frontal EEG channels. We present results suggesting that delta frequency sub-band carries most of the information associated with working memory load. Having used the above features, we also demonstrate that shorter window lengths and a smaller number of EEG channels can be successfully applied for similar contexts.