Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models