We introduce a method for learning knowledge graph embeddings as evolving Gaussian distributions, using Kalman filtering to support online updates and explicit uncertainty tracking. Our approach extends static models by replacing point embeddings (vectors) with probability distributions and applying online Kalman updates to both means and covariances. This allows embeddings to adapt continuously as new data arrives, without retraining. Experiments on standard benchmarks show up to 15% improvement in Mean Reciprocal Rank and 12% in Hit@10 over static baselines. Our method achieves scalable, real-time knowledge graph embedding with interpretable uncertainty.