Social navigation in dynamic human environments requires both efficient goal-directed behavior and robust, anticipatory safety mechanisms. We introduce BPPO-EM, a cognitively inspired reinforcement learning framework that augments Proximal Policy Optimization with Bayesian uncertainty modulation, adaptive learning rate tuning, and episodic memory recall. To enhance real-time safety, BPPO-EM incorporates a predictive collision network and a rule-based safety layer guided by LiDAR perception. In PyBullet simulations over 1000 episodes, BPPO-EM reached the goal in 82.7% of trials, compared to only 17.5% with standard PPO, and reduced the number of episodes with collisions from 60.2% to 15.8%. Moreover, the average number of collisions dropped from 2.19 to 0.50, demonstrating a 4-fold safety improvement. BPPO-EM significantly outperformed standard PPO in average rewards, returns, and goal-directed progress, confirming its effectiveness for socially compliant robot navigation.