Enhancing trajectory prediction through test-time training with a masked autoencoder and actor-specific token memory.