Reliable calibration of ecological process models is often limited by the lack of long-term observations, particularly in data-scarce forest ecosystems. Tree rings provide annually resolved growth records and represent a valuable information source for constraining process-based models. In this study, we develop a tree-ring--informed data assimilation framework to improve the parameterization and long-term simulation of forest growth using the 3-PG model. Tree-ring-derived historical biomass was assimilated into the model through Morris sensitivity screening and Bayesian MCMC calibration, substantially reducing parameter uncertainty and improving model performance across multiple structural and biomass variables. The calibrated model was applied to simulate spatiotemporal dynamics of forest net primary productivity (NPP) from 1970 to 2020 and to project future carbon density under different climate scenarios. Simulations revealed a gradual NPP decline in the study area over the past two decades. Extreme climatic events caused significant disturbances to forest growth, leading to productivity fluctuations and carbon sink instability. Forest NPP declined markedly in cold-dry and warm-dry years, while responses in warm-wet years varied. About 75% of the area was temperature-limited and 18% precipitation-limited, with temperature dominance above 2300 m and precipitation limitation below 1900 m. The current carbon stock is estimated at 77.9 Tg C, with a mean density of 112.8 t/ ha, projected to decline by up to 10% in some regions by 2060. Tree rings thus provide reliable data for model calibration and highlight increasing forest vulnerability under climate warming.