Telling a clinical voice AI to 'always finish the warning' barely helps once a patient interrupts it.

This explores whether adding protective instructions to a clinical voice agent's prompt can keep it from losing required medical content when a patient talks over it. The short answer from the corpus is: barely. In one test, four LLMs doing clinical voice tasks failed to keep required safety content in 88.6–100% of cases where the patient interrupted. How well each model recovered afterward varied from model to model with no clear pattern, and defensive prompting added very little protection Do patient interruptions break clinical voice agent safety?. Only one study in the collection tackles this exact question, so treat that as a strong signal, not a settled verdict.

The more interesting question is why a prompt fails here when small prompt changes work well elsewhere. Simply putting "wait a minute" before a question, or telling a model to correct false assumptions, sharply improved safety scores on medical-myth benchmarks Can simple discourse markers make LLMs question harmful assumptions?. Splitting a mental-health task into three explicit stages improved cognitive-distortion detection by more than ten percent Can structured prompting improve cognitive distortion detection?. The difference may be timing. Those prompts change how the model reads a single input. An interruption is something that happens over time: the agent has to remember what it hadn't said yet and come back to it after the conversation has moved on. A prompt shapes what the model pays attention to. It doesn't give the model a place to store unfinished business.

A related finding about agent memory points the same way. Agents tested on user preferences could usually recall a preference when asked but often failed to act on it, and the main cause was failing to apply stored information, not failing to retrieve it Why do LLM agents remember preferences but not act on them?. Dropped clinical content looks like the same gap: the model may 'know' the warning matters but doesn't bring it back into the conversation after an interruption. Writing that knowledge into the prompt more forcefully doesn't help if the weak point is putting it into action.

This suggests the fix belongs in the system design more than in the prompt wording. Fin's Blueprint Team argues that factual content should live in retrievable knowledge bases, and that agent instructions should hold only rules about tone and behavior. Their reason is that models follow 'what to do' rules more reliably than constraints about what to ignore or remember Should factual content and behavior rules live separately?. Applied here, required safety content could be kept as an external checklist that the system tracks and ticks off, instead of an instruction the model has to keep in mind. VoiceMem shows that voice agents can run separate memory branches and look things up during the natural pauses in speech without adding delay Can memory retrieval hide inside voice agent silence?. A similar parallel track could hold 'content still owed to the patient' and bring it back once the interruption ends.

The takeaway you might not have expected: interruption is a test of conversational memory, not of following instructions. That puts clinical voice safety closer to research on dialogue state and turn-taking than to prompt engineering. The collection doesn't yet have a study testing an external-checklist approach against defensive prompting under interruption. That gap is worth noticing.