An accurate estimation of forest aboveground biomass (AGB) is critical for quantifying carbon sequestration and supporting climate change mitigation strategies. While deep learning (DL) has advanced AGB inversion, most models rely on short-term spectral snapshots, failing to capture the decadal growth dynamics that distinguish long-term carbon accumulation from seasonal phenology. This study develops a robust AGB estimation framework using a Long Short-Term Memory (LSTM) architecture trained on a decadal (2008–2018) multi-sensor time series. The dataset bridges the transition from Landsat 7 ETM+ to Sentinel-2 MSI, incorporating 12 temporal steps across spring and summer phenological windows. To enhance model reliability, we integrated a Sensor ID embedding and a Time-Delta feature to normalize radiometric discrepancies and irregular sampling intervals. Results indicate that the LSTM framework significantly outperformed traditional Random Forest (RF) and non-temporal ablation models, achieving an R² of 0.7053. The inclusion of dual-season observations proved essential, as the 12-step model exhibited superior accuracy compared to a 6-step spring-only variant. Despite these improvements, the model demonstrated a conservative bias in high-biomass stands, with a predicted maximum AGB approximately 20 Mg/ha lower than field-measured maximums, highlighting the persistent challenge of spectral saturation in dense canopies. Our findings suggest that while decadal trajectories provide critical temporal context for biomass inversion, future improvements should focus on integrating biophysical constraints, such as growth-and-yield logic, to mitigate saturation. This research provides a scalable, geophysical consistent approach for regional biomass monitoring in complex temperate and boreal forest ecosystems.