Currently, my agent with speech-to-speech voice (OpenAI Realtime) is working completely decoupled from my chat's memory. How can I have my voice transcriptions in my thread's memory and streamed to the client? I have a SvelteKit client with working chat interaction.