With the growing number of satellites in Earth orbit, inter-satellite navigation in Medium Earth Orbit (MEO) faces increasing challenges due to high noise levels and enhanced system nonlinearity. Traditional filtering methods often struggle to maintain accuracy and stability in such complex environments. To enable precise adaptive tuning of the process noise covariance, this paper proposes a Q-learning Extended Kalman Filter with a pruning-and-growth state-space mechanism (QLEKF-Cut). Built upon the QLEKF framework, the proposed method introduces an adaptive state management strategy that periodically prunes states with low average rewards to improve computational efficiency, while dynamically expanding the state space when high-value states lie at the boundary, allowing for fine-grained exploration in critical regions. Simulation results demonstrate that this method effectively avoids inefficient exploration and significantly improves both navigation accuracy and system robustness, making it well-suited for MEO inter-satellite navigation under complex noise conditions.