Thanks — I’m glad you asked. I’ll answer both parts of your question: whether system instructions stay constant, and whether LiveKit Inference performs any caching. • System instructions stay constant for an agent session. The

instructions

you pass to an agent define its behavior and remain in effect for the session unless you explicitly update them (for example, via an update function or API call). • LiveKit Inference does not cache LLM inference results. Each STT/LLM/TTS request is routed to the model provider and processed in real time; there’s no built-in memoization of inference responses. That means you should expect fresh model processing for each turn unless you implement your own caching layer. • There is separate caching support for TTS audio only. You can pre-synthesize or cache audio frames for repeated text to avoid re-synthesizing the same audio, but that caching is specific to audio output and does not affect LLM inference or system instruction behavior. If you’re planning to rely on a constant instruction but want to optimize repeat responses, tell me more about your use case (e.g., predictable prompts, repeated short phrases, or high-throughput sessions) and I can suggest practical caching strategies or where to apply TTS caching without altering inference consistency. Sources: Build a salary negotiation coach with a LemonSlice avatar | LiveKit | Agent speech and audio | LiveKit Documentation | Introducing LiveKit Inference: A unified model interface for voice AI | Liv..., with interactive elements