Accurate volatility forecasting is essential for financial risk management in the digital asset sector. However, traditional econometric models often struggle to capture the non-linear dynamics and heavy-tailed distributions of cryptocurrency returns. Furthermore, existing literature frequently relies on an assumption of temporal uniformity, ignoring the distinct market microstructures of different assets. To address these limitations, this study proposes a novel hybrid approach that integrates GARCH-family models (Standard GARCH, EGARCH, and GJR-GARCH) with a Bidirectional LSTM (BiLSTM) network. We employ a data-driven grid search strategy to identify the optimal window size for specific assets. Additionally, we integrate the Quasi-Likelihood (QLIKE) loss function to ensure robustness against extreme outliers. Our empirical results reveal divergent memory regimes: Bitcoin minimizes forecasting error at a 60-day window (reflecting macro-driven long memory), while Ethereum optimizes at a 21-day window (reflecting utility-driven ecosystem cycles). By incorporating Explainable AI (XAI) principles via saliency mapping, we visually validate these distinct temporal dependencies. The proposed model significantly outperforms the best econometric baselines, reducing the Mean Absolute Error (MAE) by 9.4% for Bitcoin and 21.4% for Ethereum. Consequently, this study demonstrates that asset-adaptive forecasting not only improves statistical precision but also enhances capital efficiency for institutional risk management frameworks.