Variability in brain signals due to factors such as inter-individual differences and session-specific shifts limits the utility of Brain-Machine Interfaces (BMI) outside controlled environments, particularly in motor rehabilitation. This study examines the use of Riemannian processing to represent manifolds in electroencephalography (EEG) signals, providing a more consistent and plastic representation of brain activity during motor imagery tasks. We analyze a publicly available dataset comprising five motor imagery training sessions involving nine subjects. Our approach involves projecting covariance matrices of EEG epochs into the tangential Euclidean space and reducing dimensionality using Principal Component Analysis (PCA). The Riemannian processing is compared with Common Spatial Patterns (CSP) to assess the invariance and discrimination properties of the resulting data distributions. Results indicate that Riemannian processing successfully centers the data and aligns distributions across sessions and subjects while maintaining class discriminability. Furthermore, we observe manifold changes over practice, with increased distance between motor imagery classes reflecting improved learning. These findings suggest that it is possible to represent latent manifolds in non-invasive EEG data, providing a basis for BMIs that generalize better across individuals and sessions without recalibration.