How I built a zero-token memory layer for LLMs (and why it outperforms vector store approaches)