The patient advocate
Being seen and heard is part of getting better. A physician with forty-five minutes to divide among a dozen patients cannot also assemble the whole story out of unfamiliar diagnoses, new prescriptions, and pages of lab results, and neither can the patient sitting in the same room. The appointment is short and it comes around rarely, so most of what decides how it goes has to happen before it starts.
Patient advocates step into this space as a second set of eyes and ears. They provide a second mind that recalls conversations from months prior, connects scattered data points, and independently researches the gaps left unexplained during a brief clinic visit. Most people go through this alone, so the question worth working is what it takes to build that second mind as an agent.
the anatomy of a second mind
Each of the advocate’s duties lands on a specific piece of agent architecture.
Recall is long-term memory. Following a session, the system distills the transcript into structured records, each carrying its date, its source, its exact values with units, and its current status: [FACT | 2026-07-03 | src: discharge orders | active] metoprolol tartrate 25 mg twice daily. These records are embedded as vectors and securely stored, and the structure lets the record evolve. A correction from the patient supersedes the old entry, retiring it to inspectable history rather than deleting it. When a new question arises, the agent retrieves the most clinically relevant history through three channels at once: semantic similarity, the index keys a record is filed under, and the dates it is about.
Connection comes out of that accumulation. A creatinine level reads one way on its own and another way as the third consecutive rise, and only a stored history puts the second reading within reach.
Research goes to a specialized subagent with access to current, fully sourced clinical information. A language model’s training data is frozen and unattributed, which makes it the wrong place to look up a dose.
the two jobs, worked through
This architecture does two jobs for the patient.
First, it explains symptoms and diagnoses. If a patient receives a prescription for metoprolol following a heart event, the advocate clearly translates the function of a beta-blocker. It separates expected side effects from genuine emergencies and puts the critical safety rules first, such as the danger of stopping the drug suddenly. The numbers themselves stay where they belong: any hold-the-dose cutoff or when-to-call threshold appears only as a quote from this patient’s own record, attributed to the clinician who set it, with its date. When no such instruction exists in the record, the advocate says so and drafts the exact question that gets one.
Second, it prepares the patient for follow-up care. The system drafts precise questions for the coming appointment, so the physician’s limited minutes go to the questions that matter. Here is that process in a live cold-started session; watch the recalled records arrive with their sources attached, and watch a correction supersede the old dose:
The operational boundaries are written directly into the prompt. A prompt is instruction, not enforcement; the model can still fail these lines, which is why the emergency override is stated first and the line on numbers leaves nothing for the model to invent:
- EMERGENCY OVERRIDE: If symptoms indicate a potential emergency, the FIRST sentence must direct the user to call 911 or seek immediate care.
- YOU PREPARE, THEY PRESCRIBE: Never instruct the user to start, stop, or alter a medication. Provide the exact questions for the clinician who holds that authority.
- NUMBERS BELONG TO THEIR CLINICIAN: Numeric thresholds appear only as quotes from this patient's record, attributed and dated. An invented threshold is a prescription in disguise.
- Maintain strict honesty regarding absent records; never invent a remembered fact.A properly designed second mind strengthens the physician-patient relationship. It improves the conversation instead of offering a competing medical opinion.
build it yourself
The complete system prompt is yours, with its identity, memory doctrine, communication laws, research rules, safety boundaries, and deployment notes: download the full prompt (markdown). It assumes only a persistent fact store, a search tool, and a date binding; the reference implementation runs on melchizedek’s long-term tier — config, memory pipeline, and schema all in the public repo. One non-negotiable: memory siloed per patient and deletable by the patient. A record this intimate belongs to the person it describes, or the agent shouldn’t exist.
Being heard, it turns out, has an architecture. It is someone who remembers what you said, connects what you couldn’t, and walks into the room already holding your questions. Most people never get that someone. Now you know how to build one.