Traditional assessment instruments for post-stroke aphasia are typically extensive and asynchronous, providing results only after full administration. This delay limits their suitability for continuous clinical monitoring and real-time adaptive therapeutic adjustments. Objective: This study evaluates the feasibility of a low-cost, low-density EEG-based framework designed for the continuous estimation of working memory capacity as a key functional biomarker of rehabilitation progress. Methods: Electroencephalography (EEG) data were acquired from aphasic and control participants using a portable OpenBCI platform and a customized 3D-printed helmet of an n-back paradigm. Event-related potential (ERP) features, including P100, N100, P200, N200, and P300, were extracted to train linear (OLS) and a non-linear Support Vector Regression (SVR) model with a radial basis function (RBF) kernel against a composite ground-truth index (W). Results: Non-linear SVR models demonstrated superior performance and scalability compared to linear approaches. For groupings of 32 epochs, a statistically significant improvement was identified (p=0.04), with non-linear models showing a reduction in RMSE up to six times greater than linear models as effective training samples increased. Conclusion: These findings provide a proof-of-concept for a deployable, technology-assisted system capable of supporting individualized and adaptive cognitive monitoring in clinical post-stroke rehabilitation settings.