Brain-Computer Interfaces (BCIs) based on Speech Imagery (SI) hold transformative potential for restoring communication in individuals with locked-in syndrome, Amyotrophic Lateral Sclerosis (ALS), and severe dysarthria. However, practical deployment of Speech Imagery BCIs is critically limited by the cross-session transfer problem: Electroencephalography (EEG) signal distributions shift substantially between recording sessions due to electrode repositioning, impedance fluctuations, and cognitive state variations, causing within-session classification accuracies of 60–70% to collapse to 30–40% in cross-session evaluation. Existing deep learning architectures for EEG decoding are hand-designed and optimized for within-session performance, leaving the cross-session generalization problem largely unaddressed. We propose Metaheuristic-guided Neural Architecture Search (NAS) for Speech Imagery (MetaNAS-SI), a novel framework that automatically discovers optimal neural network architectures explicitly optimized for cross-session EEG Speech Imagery decoding. MetaNAS-SI introduces a hybrid evolutionary search combining aging evolution for macro-architecture exploration with Differential Evolution (DE) for micro-architecture optimization, a cross-session fitness function based on Leave-One-Session-Out (LOSO) evaluation, a searchable Session-Invariant Feature Alignment Module (SIFAM) that performs domain adaptation via maximum mean discrepancy within candidate architectures, and an EEG-specific search space incorporating physiologically motivated temporal convolutions, Common Spatial Pattern (CSP)-inspired spatial filters, and multi-scale temporal attention. MetaNAS-SI was evaluated on three Speech Imagery datasets: KaraOne (5-class), the coretto imagined speech dataset (4-class), and an in-house Farsi-Arabic dataset (6-class, 5 sessions per subject). Cross-session accuracy reached 45.8% on KaraOne (vs. 36.4% for the best baseline), 52.3% on Coretto (vs. 43.8%), and 48.1% on the Farsi-Arabic dataset (vs. 35.2%). Discovered architectures were 3–15× smaller than hand-designed alternatives while achieving 18–38% relative improvement in cross-session accuracy. Ablation analysis confirmed the critical contributions of the cross-session fitness function and SIFAM module. Pareto-optimal architecture selection enables deployment-aware trade-offs between accuracy, model size, and inference latency for real-time BCI applications.