A KISOP field guide

Keeping Your Data Safe with AI Tools

Between last October and this May, every large U.S. technology company introduced a way to bring health data into its products. This guide explains what each product does well, what to check before connecting it, and one protective habit: leave every service with copies.

This page is updated continuously; last checked September 9, 2026. If you have a recommendation or a correction, send it through Write to us → feedback.

What just happened, and who built the tools

Each company built the same thing: a dedicated space where you connect your medical records, your labs, your wearable, and an assistant reads them back to you. They even share the plumbing. HealthEx, a patient-consent platform riding TEFCA's individual-access rails, delivers your records to both Claude and Copilot. b.well does the same job for ChatGPT and Google. Amazon taps state health information exchanges directly. These are all built on your federal right of access, which has existed since the 21st Century Cures Act made information blocking illegal.

So everyone now has access to your records, and it is your access they are borrowing: every one of these connectors works by exercising, on your behalf, the legal right you always had and almost never used. The competition is over the relationship: which assistant remembers you, which app becomes the place where your health seems to live. Memory is the moat they are digging. The only question that matters on this page is whose memory it turns out to be — theirs, or yours.

I spent years at Google Health building systems that pulled in enormous numbers of health data sources, and the lesson never left me: convert-everything pipelines are brittle, expensive, and clinically risky — and anything you pour yourself into can be switched off. In January 2023 it was. That is not a grievance; it is a design requirement.From "Health You Own," the essay behind this site

Two rules before any playbook

Rule one

Share your data in the right ways when you use a Chatbot or AI-enhanced tools

At KISOP, we want to help you keep your health information private and secure. Systems and laws change all the time. Presently (August 2026), most systems offer two ways for you to interact side by side. The dedicated health spaces (ChatGPT Health, Claude's health connectors, Copilot Health) are ring-fenced: all four companies state that what happens there is not used to train their models, and the record connections are permissioned category by category.

The regular chat window of the very same product has a different set of rules. Claude's consumer chats have trained models by default since September 2025 unless you switch it off (Settings → Privacy), and saying yes extends retention from 30 days to five years. Anthropic updated its consumer policy again on July 8, 2026 (multi-step tasks, connected apps, identity verification), and a separate Consumer Health Data Privacy Policy has applied to connected health apps since January 12, 2026. The Terms & Conditions of any one company change all the time. We encourage you to read them as they are shared so that you can continue to use your data safely. ChatGPT's general window trains unless "Improve the model for everyone" is off. Paste your labs into the wrong tab and you have made a donation you didn't intend.

Under HIPAA, companies are required to sign Business Associates Agreements that govern the care provider (doctor) and insurer side of the business if your Protected Health Information (PHI) is at risk of being exposed. Learn about HIPAA here: https://www.hhs.gov/hipaa/index.html. When you connect your own records to a consumer app under your right of access, Health & Human Services treats the app as your choice, not the provider's business associate, so HIPAA stops at the hand-off. The rules of the Federal Trade Commission (FTC) and the app's own terms of service cover it from there. The privacy you have in these spaces is the privacy the company chose. So check it yourself, and write down the date you checked.

Rule two

Plan the exit on the day you enter

All of these companies are trying to provide a useful service. Even great services get shut down for reasons we can't control. KISOP is trying to build a world where you can use these services when they serve you, and also have choices in case these services become unavailable. Below are examples:

The graveyard
PlatformLifespanHow it ended
Google Health (v1)2008–2012Retired for low adoption; users given a window to download, then gone.
Microsoft HealthVault2007–2019Shut down November 20, 2019. Remaining data deleted; every app built on it stopped working.
Amazon Halo2020–2023Discontinued with refunds; health data erased on schedule.
Apple's AI coachshelved Feb 2026Cancelled before it ever shipped.
The Fitbit apprenamed May 2026Absorbed. Your login survived. The product you chose is gone.

Microsoft, whose new Copilot Health asks for your medical records, is the same company that deleted its last health vault six years ago. Use these tools expecting them to end. If your copies outlive their product decisions, you lose nothing when the memo goes out. If they don't, you start over from zero, again, at whatever age you are when it happens.

The playbooks

For each environment: its strongest use, setup steps, risks to watch, and how it can work with an archive you own.

Claude — Anthropic

Claude for Healthcare, January 2026 · records via HealthEx & Function connectors (beta) · Apple Health / Health Connect in beta · requires Pro or Max

Best for: deep reading. Claude's health integration is built on a retrieval protocol (MCP) that pulls only the categories relevant to your question (medications, allergies, recent labs), and Anthropic states connected health data isn't retained in memory or used for training. Architecturally, it is the most conservative of the five. In practice Claude is the assistant I hand forty pages of visit notes and ask what the three specialists missed between them. Long documents, ambiguity, competing evidence: this is its terrain.

Do this

  • If you subscribe: connect HealthEx once, grant only the categories you need (conditions, labs, medications, notes), and run the pre-visit brief before each appointment (prompt at the bottom of this page).
  • If you don't subscribe, paste the relevant slice from your own files into a chat — a season of labs or a medication list. Keeping that source slice in your archive makes the subscription optional.
  • Set up a dedicated Project as your health workroom, with your baseline (conditions, meds, allergies, the values that run abnormal-for-you) in the project instructions, so every conversation starts already knowing your thousand variables.
  • Export your conversations periodically (Settings → Privacy → Export data). The reasoning you and Claude build over months is itself health information. It belongs in your archive.
WatchIn the general chat window, training on your conversations is on by default (since September 2025) with five-year retention. Flip it off in Settings → Privacy if that's not a donation you mean to make. And the health connectors are beta features of a subscription product: priced, versioned, and cancelable like any other.
With your archiveClaude is the interpreter. The shelf is yours. Feed it slices; when it produces something worth keeping (the brief, the timeline, the "ask your cardiologist these five things"), save that output into your archive as a dated file, labeled AI-generated, Claude, July 2026. Interpretations age; label the vintage.

ChatGPT — OpenAI

Health in ChatGPT — early access January 2026; open to every U.S. adult on the Free, Go, Plus, and Pro plans since July 23, 2026 · records via b.well, One Medical, and Function · Apple Health, MyFitnessPal, Weight Watchers connectors

Best use: an ongoing conversation. ChatGPT Health has separate memory for its files, chats, and connections; OpenAI states that this material does not enter regular ChatGPT memory or train its foundation models. Its strength is continuity: tracking symptoms across weeks, discussing medications in context, and returning to follow-up questions later.

Do this

  • Use the Health space for anything with your name or your numbers in it. The room exists; be in the room.
  • Connect records through b.well selectively: start with labs and medications, the two categories with the highest question-per-page yield.
  • In the general window, turn off "Improve the model for everyone" (Settings → Data Controls), because one day you'll forget which tab you're in.
  • Export your data (Settings → Data Controls → Export) on a calendar rhythm (quarterly is enough) and file the health-relevant conversations into your archive.
  • Treat its guidance as preparation: OpenAI's own framing is that it isn't for diagnosis or treatment. Use it to arrive at appointments sharper. Keep the appointments.
WatchThe Health space opened to everyone in July 2026, and the walls are policy: "stored separately" and "not used for training" are commitments in a document the company can revise. Read the data controls again each time the product changes name, price, or owner of the moment.
With your archiveAnything stored only in ChatGPT Health remains inside another company's service. When a thread produces a useful insight — a migraine pattern or a question that led to a diagnosis — summarize it, date it, and file it. A future assistant will not inherit that history unless you carry it forward.

Google Health — Google

The Fitbit app, renamed May 19, 2026 · Gemini-powered Health Coach in Premium ($9.99/mo) · records via b.well + CLEAR identity · Fitbit & Pixel Watch native, Apple Health importable

Best use: For Fitbit or Pixel Watch users, Google Health can place sleep, heart rate, and steps beside labs and visit history. That combination may reveal patterns a clinic portal does not present. Its records connection — through a provider portal search or CLEAR identity verification — is the simplest onboarding among the five systems reviewed here.

Do this

  • If you're in the Fitbit/Pixel world: link records, try Coach for one quarter, and judge it on one number you're trying to change.
  • Run Google Takeout twice a year: it exports your complete Fitbit history (years of sleep, heart rate, activity) as CSV/JSON into your archive. This is the single most valuable export in consumer health and almost nobody runs it.
  • On Android, treat Health Connect as your hub: it's the on-device layer where all your apps' health data crosses, permissioned per app and per data type. Audit those grants quarterly.
  • Before believing a coach insight, ask it what data it used. If the answer is two weeks of steps, weigh it accordingly.
With your archiveYour archive's Device-Data partition is the permanent home; Google Health is the live dashboard. Takeout feeds the archive; the archive means that when you switch watches (or Google switches strategies), fifteen years of your body's rhythms move with you. Without that archive you start again at zero.

Copilot — Microsoft

Copilot Health — announced March 12, 2026; in open preview since May 29, 2026 for U.S. adults with a Microsoft 365 Personal, Family, or Premium subscription · records from 50,000+ provider orgs via HealthEx · Apple Health connected, more wearables promised · Microsoft states the health space is not used to train its AI

Best for: the coherent story. Microsoft's pitch is aggregation: records plus wearables plus labs, assembled into one narrative of your health, from the company already answering fifty million health questions a day across Bing and Copilot. If your life runs on Windows and Microsoft 365, Copilot Health will be the ambient option, and its clinical-grounding work (Harvard Health content licensing, careful sourcing) is real.

Do this

  • Try the preview if you're curious, but connect selectively on day one: records yes, wearables after you've seen how it handles the records.
  • Ask it for the "coherent story" and then check the story against your own files: early aggregation products earn trust by being audited.
  • Before connecting anything, locate the export and the disconnect. If you can't find both in five minutes, wait a version.
WatchHealthVault. Microsoft ran the "trust us as your health vault" experiment from 2007 to 2019, then deleted the vault and orphaned every app built on it. The people are different, the architecture is better, the incentive structure is identical. Microsoft also frames Copilot Health as sitting outside HIPAA at launch. Their candor is useful; borrow it.
With your archiveIf Copilot creates a useful medication timeline or condition summary, export it to PDF, date it, and file it. Aggregations can be regenerated when you retain the source records.

Apple Health — Apple

The Health app + Health Records (FHIR from hundreds of health systems) · AI coach shelved Feb 2026; a redesigned Health app arrives with iOS 27 (September 2026), the coach still absent · on-device, encrypted

Best use: device-centered custody. Its Health app stores data on the device, supports end-to-end encryption, and combines watch, phone, and provider FHIR records without advertising. It currently lacks an AI interpreter and remains tied to the iPhone.

Do this

  • If you're an iPhone person, turn on Health Records for every provider that supports it: it's the lowest-effort continuous records feed a consumer can get in the U.S.
  • Twice a year: Profile → Export All Health Data. You get a ZIP of everything: sensor data as XML, clinical records as FHIR JSON. File it in your archive even though the XML is ugly; ugly and yours beats elegant and theirs.
  • Use the connectors: Apple Health is what Claude, ChatGPT, and Google all read from on iOS, which makes it your staging layer, the place where permissions get decided app by app.
  • Export now, without waiting for Apple's coach, and bring the data to whichever interpreter you've chosen this year.
WatchThe gravity is the device. Everything works beautifully until the year you want an Android, a different watch, or a family member on another platform, and the coach that was coming "next year" has now been shelved once already. Apple's caution is to its credit. Build your health strategy on your own timeline.
With your archiveClosest cousin, still a silo. Treat Apple Health as the best-behaved source your archive has: it gathers beautifully, exports completely (if clumsily), and asks little in return. The export ZIP → your archive → any AI on any platform.

Amazon — One Medical & Health AI

Health AI on Amazon.com and the app, free for every U.S. customer with no Prime or One Medical membership required (March 2026) · reads records via state health information exchanges · One Medical primary care · 5 free virtual visits for Prime

Best for: collapsing the distance between question and care. Amazon's assistant is the only one that can read your history, answer the question, book the appointment, and refill the prescription, because Amazon is the only one that also employs the doctors and runs the pharmacy. For the uninsured and the just-get-it-done, an agent with five free virtual visits attached is access. This is health care delivered through Amazon's logistics network.

Do this

  • Use it for refills, common-condition visits, and "do I need to see someone for this" triage, where a shorter wait can improve care.
  • Know which Amazon you're talking to. One Medical, the provider, is inside HIPAA; the Health AI assistant on Amazon.com is a consumer product governed by its own terms. Same brand, different laws.
  • If you use One Medical: request your full record (they must provide it: right of access) at least yearly, and file it. Care records from a convenience-first provider are still care records.
  • Consider your purchase history and health questions together, because Amazon may be able to connect them. Decide whether you accept that proximity.
WatchHalo. Amazon entered body-data with hardware in 2020 and erased the category (and the data) in 2023. And the deeper geometry: you are a customer of a store that now knows your diagnoses. The incentive is commercial, which is exactly why it needs your eyes open.
With your archiveAmazon optimizes the transaction; your archive holds the continuity the transactions never see. After each One Medical visit, the summary goes in your archive next to the labs from Quest and the notes from Stanford, because you are the only integration layer those three will ever share.

Verily Health — Verily Me & Lightpath

Verily Me, free, October 2025 · AI agent "Violet" + licensed clinicians who review records across systems · Lightpath cardiometabolic programs via employers/payers · independent from Alphabet and renamed Verily Health, April 2026

Best use: licensed clinicians in the loop. Verily, formerly an Alphabet company and now Health Inc. (following a completed $300 million funding round that transitioned the business into an independent corporation), is the only offer on this page that includes clinician review by default. Verily Me is free: Violet answers questions about your data, clinicians review records across providers, and meal photos receive nutrition feedback. Lightpath adds a care team for diabetes, prediabetes, and weight when an employer or plan pays.

Do this

  • If your employer or plan offers Lightpath and you're managing a cardiometabolic condition: evaluate it seriously. A care team plus a 24/7 companion is a different category from a chatbot with your labs.
  • In Verily Me, use the clinician review. A licensed human reviewing cross-system records at no charge is unusual; request a copy of what they assemble.
  • The Lifelong Health Study invitation is a real decision: research participation is honorable (I believe in consented research), but read the consent as carefully as you'd want researchers to read your data. Know the withdrawal terms before you enter.
WatchCorporate mobility. Verily has been Google X, then Alphabet's bet, now Verily Health, an independent company since April 2026, raising to fund an AI roadmap: three identities in a decade. Programs subsidized by employers end when contracts end. Enjoy the clinical rigor; keep your copies current.
With your archiveVerily models the KISOP conviction that people need people: data doesn't explain itself. An assistant that can escalate to a human is the right shape. Your archive is what makes you a better participant: you arrive with your history organized, and everything their clinicians produce (reviews, care plans, coaching notes) flows back into it, labeled and dated.

Muse — Meta

Muse, September 8, 2026 · a personal agent that does tasks for you: email, bookings, purchases, long-term plans · asks for access to email, calendar, financial, health, fitness, and smart-home apps · iOS, Android, muse.ai, and WhatsApp; AI glasses to follow · free for most use, subscriptions above that · U.S. first

Best for: nothing health-related yet. The launch pages never mention health, and the app asks for your health apps anyway. What Meta describes is an agent that reads your accounts, makes a plan, and acts, inside what it calls a Secure VM: a dedicated computer Meta runs for you, with a second program, Sentinel, watching what leaves it. Meta calls the direction "personal superintelligence." Read that as a place that wants your whole life, health included, before it has said what it will do with the health part.

Do this

WatchMeta promises a Confidential VM "later this year," encrypted with a key only you hold, "so not even Meta can access it." That version does not exist yet. Of the version that does, Meta's VP of Engineering, David Singleton, said policy keeps Meta out of the Secure VM and that Meta technically could reach in, as quoted by SiliconANGLE. The model is Meta's own Muse Spark; no open weights are mentioned. Meta's record on personal data includes the 2019 Federal Trade Commission order and its $5 billion penalty.
With your archiveThe Confidential VM, if it ships as described, is the closest a large company has come to the idea behind this site: a machine holding your life that only you can unlock. Read the difference carefully. Muse keeps your life on a machine Meta owns, under terms Meta writes, in a product Meta can end. The Personal Neutral Archive (PNA) is a set of files you hold, on a device you choose, that no company can switch off. If you use Muse, use it the way this page says to use every rented assistant: hand it slices, keep the originals, and save what it produces into your archive, dated and labeled AI-generated, Muse, September 2026.

Open models — the AI you run yourself

Open-weight models from Google (MedGemma 1.5, January 13, 2026), NVIDIA (Nemotron, Clara), Meta, Alibaba, Mistral, and OpenAI (gpt-oss) · found on Hugging Face · run on your own computer with Ollama, LM Studio, AnythingLLM, or NVIDIA's ChatRTX · no account, and no terms of service on the model once it is on your disk

Best for: reading your own files where they already live. An open model is a file you download once and run on your own machine. Nothing you type leaves the device. There is no subscription to lapse and no policy page to re-read, because the model cannot phone home. MedGemma is the medical example: Google trains it, publishes the weights on Hugging Face under its Health AI Developer Foundations terms, and the 1.5 release reads CT and MRI as well as text. The general models read a lab report or a visit summary well, and the small ones run on a laptop. NVIDIA matters here twice: its GPUs are what most of these models run on, and it now publishes open models and a local runner (ChatRTX) for its RTX computers, so a desktop machine can hold a private assistant with no cloud behind it.

Do this

WatchOpen models are built for developers first. There is no support line and no clinician review, and the capable ones want a strong computer with a good graphics card. Hugging Face is a company too, and licenses differ model by model; the model file, once on your disk, stays yours regardless. A small local model reads less well than the large rented ones today. For many questions it is enough; for the hardest ones, use both.
With your archiveThis is the option the rest of this page points toward. Your archive holds the files; a local model reads them without the files leaving the room. KISOP's design keeps a local-AI-only mode for the same reason some clinicians decline cloud AI on patient data. The rented assistants stay useful, and you can use them when they serve you; the local one is the choice you keep in reserve.

Side by side

The facts that decide what to connect, as of July 2026, re-checked September 2026. Scroll sideways on a phone.

EnvironmentThe health offerHow records arriveTrains on health space?General chat trains?What you can carry outTrack record
ClaudeHealth connectors in Pro/Max; per-question retrieval (MCP)HealthEx, Function; Apple Health / Health Connect betaNo, per Anthropic; not retained in memoryYes by default since Sept 2025: opt out in Settings → PrivacyConversation export; your connected sources remain yoursNo prior consumer health product
ChatGPTChatGPT Health — sealed space, own memoryb.well; Apple Health, MyFitnessPal, WWNo, per OpenAIYes unless "Improve the model" is offFull data export (ZIP)First health product; open to all U.S. adults since July 2026; young walls
Google HealthApp (ex-Fitbit) + Gemini Coach, $9.99/mo Premiumb.well + CLEAR identity searchRecords "under your control," per GoogleGemini Apps Activity — review settingsTakeout: full Fitbit/device history, CSV/JSONGoogle Health v1 retired 2012; Fitbit brand retired 2026
CopilotCopilot Health — records + wearables narrativeHealthEx (50k+ provider orgs); Oura, FitbitIsolation commitments statedCopilot personalization — review settingsTBD — locate export before connectingHealthVault deleted 2019, apps orphaned
AppleHealth app + Health Records; coach shelvedDirect FHIR from providers; on-deviceOn-device; no assistant reading it yet—Export All Health Data: XML + FHIR JSON ZIPHealth app steady since 2014; AI plans cut Feb 2026
AmazonHealth AI agent + One Medical care + pharmacyState health information exchangesConsumer terms govern the assistant—Right-of-access request to One Medical (HIPAA side)Halo shut down 2023, data erased
Verily HealthVerily Me free + Violet + clinician review; Lightpath programsCross-system record review; you link portalsResearch only with study consent (your choice)—Request the clinician-assembled record copySerious clinical culture; three corporate identities in a decade
Muse (Meta)No health offer named; the agent asks for health and fitness appsYour connected accounts, read from a machine Meta operatesNot statedOpt-out offered; the default is not stated. Check on day oneNot stated; find the export before connectingFTC order and $5B penalty, 2019; Confidential VM promised for late 2026
Open models (run yourself)MedGemma, Nemotron, Llama, Qwen, gpt-oss and others, via Hugging Face; run with Ollama, LM Studio, ChatRTXYour own files, on your own machineNo: nothing leaves the device—Everything; the model file is yours once downloadedDeveloper tools; no clinician; the model stays even when a company changes course

Read the last column twice. Every row but the last is a room in someone else's building. The last row runs in your own house, and it still needs a place to keep what you feed it and what it produces. A place that is yours is what all nine are missing: the archive.

The missing piece: a reservoir of your own

Every platform above is a faucet. The reservoir is the one thing none of them can sell you.

The tandem model is almost embarrassingly simple. You maintain one canonical archive of your health information: plain files, in the formats they arrived in, on storage you control, and encrypted: sensitive data should be safe where it lives, and the weekend list below says how. A folder tree. Labs as the PDFs and CSVs the portal gave you. Records as the FHIR files Apple exported. Your Fitbit decade as Takeout delivered it. Imaging as the DICOM disc the radiology desk handed over, because you asked. Each item with a line of provenance: where it came from, when, how. And then (this is the part people miss) every AI output worth keeping goes back into the archive too, in its own partition, labeled as AI-generated with the model and the date. The interpretations are data. Keep them, date them, and file them apart from the originals.

The archive is canonical. Every index is disposable.

That one rule is the entire architecture. Anything derived (a platform's summary, a coach's dashboard, an assistant's memory of you) can be rebuilt from your archive at any time. Your archive can be rebuilt from nothing else. A system that violates this (where the useful truth lives in somebody's database and your files are the afterthought) is a product with your data in it. A system that honors it is your archive, with products attached. You can test any tool on this page against that sentence in thirty seconds: if it ended tomorrow, what would I still hold?

The tandem loop Your rights deliver records into your archive; you bring slices of the archive to any AI assistant; distilled, labeled outputs return to the archive. Assistants come and go; the archive persists. Claude ChatGPT Google Health Copilot whatevercomes next THE FAUCETS — INTERPRETERS, RENTED questions + scoped slices outputs, labeled + dated Your Personal Neutral Archive Plain files · original formats · provenance on every item Labs · records · imaging · device data · notes · AI outputs (labeled) On storage you control — consent decided grant by grant YOUR LEGAL RIGHT OF ACCESS portals · exports · Takeout · records requests Assistants come and go. This stays — and compounds.
The tandem loop: rights fill the reservoir; faucets interpret it; distilled outputs return, labeled. Any faucet can be replaced without losing a drop.

Why does the tandem matter so much right now? Because of what these five launches actually changed. When you connect records to an assistant, the assistant gets a window, and you get answers, which is a fair trade. But the same rails that fill their window can fill your reservoir: the portal that HealthEx reads from will hand you the download; the Takeout export contains more of your Fitbit history than the coach will ever surface; the right of access that b.well exercises in your name works just as well when you exercise it for yourself. Connect if it serves you, but download first, and download always. The connection may help today. Your copies remain useful for decades.

I call the destination a Personal Neutral Archive. Neutral matters in five ways:

Vendor-neutralAny program that opens plain files can read it. Leave any tool you like and the archive comes with you.
Format-neutralOriginals kept as they arrived (FHIR, DICOM, CSV, PDF). No master schema is forced on them. Preserve at write time; interpret at read time.
Model-neutralAny AI you trust can be granted scoped, logged access; the most sensitive partitions can be restricted to models that run locally, on your machine. Every successor model inherits the full richness.
Environment-neutralPhone, laptop, home server, a cloud you choose — it's a directory tree. It copies, syncs, and replicates anywhere.
Purpose-neutralClinic data is one part. Health also includes food, housing, work, finances, and a parent's care file: one archive, partitioned, with a sensitivity level for each section.

This is also how your own AI finally reaches parity with institutions. Hospitals and platforms work from processed summaries of you: the radiologist's report stands in for the image, the discharge summary for the stay. An archive that keeps raw and processed together, with the source schemas, means any model you choose sees as much of you as any institution's environment ever has. Their view of you is a slice. Yours is the whole, and it gets richer every year. Every assistant you ever hire inherits it on day one.

Start this weekend

Two hours, no new products, nothing to buy. In order:

  1. Make the tree. One folder, named Archive, wherever you control storage. Inside it: Medical-Records, Bloodwork, Medications, Imaging, Device-Data, Notes, AI-Outputs. Done. That structure will outlive every product on this page.
  2. Lock it. Anything stored on your own machine should be encrypted, because sensitive data should be safe where it lives. Turn on full-disk encryption before you fill the folder: FileVault on a Mac, Device Encryption or BitLocker on Windows; on a phone the storage is already encrypted once you set a strong passcode. If the Archive lives on an external drive or a USB stick, encrypt that drive too (both systems can), and keep the recovery key somewhere other than the drive itself. Any copy that travels, in a backup, a shared folder, or a message, should be sealed before it leaves.
  3. Drain the portals. Log into each patient portal you have (MyChart and its cousins) and use Download my record / Export. Save exactly what they give you (PDF, XML, FHIR) into Medical-Records. Ask the lab portals (Quest, Labcorp) for the same. This is a legal right: since the Cures Act, refusing or slow-walking you is information blocking.
  4. Empty your phone's pocket. iPhone: Health app → profile → Export All Health Data → the ZIP goes to Device-Data. Android/Fitbit: Google Takeout → Fitbit + Health Connect. Ten years of your body's signals, suddenly yours in bulk.
  5. Write the provenance line. In each folder, a plain-text note: what this is, where it came from, the date you pulled it. Sixty seconds per file now; the difference between an archive and a junk drawer in five years.
  6. Now, and only now, pick one faucet. Choose one platform from the playbooks above, connect or paste deliberately, and run the brief below before your next appointment. Feel what it's like when the AI works for your archive. Every platform above makes your data work for its model.
  7. Close the loop. Whatever the assistant produces that's worth keeping goes into AI-Outputs, labeled with the model and the date. You just ran the whole tandem, once. Now it's a rhythm.
The pre-visit brief — paste this into your chosen assistant, with your records connected or your files attached I'm preparing for an appointment with [specialty] on [date]. From my records: (1) List my active conditions and current medications with doses. (2) Pull the last 12 months of labs relevant to this visit and note the trends. (3) Flag anything in my history this clinician should know before doing anything routine — allergies, adverse reactions, and values that run abnormal for me even if they look borderline-normal on paper. (4) Give me five questions worth asking. One page. Mark anything you're inferring rather than reading directly from a record.

Point 3 is the whole reason this page exists. I once told a specialist, before a routine procedure, that my electrolytes run off and my magnesium needed watching. It wasn't in her summary of me, so it didn't exist. I ended up in the emergency room over about thirty cents of magnesium. The brief puts your thousand variables on one page nobody can skip past. That is what all of this (the rails, the faucets, the reservoir) is actually for.

Where this goes

Everything on this page you can do alone, this weekend, with a folder and your legal rights. But you shouldn't have to do it alone, and most people never will: the exports are clunky, the portals are hostile, data doesn't explain itself, and AI doesn't love anyone. That's the work KISOP exists to do: make the archive effortless to keep, put a consent gate in front of it so every access is asked-and-logged, fund capabilities as community jobs-to-be-done so they belong to everyone once built, and put Community Helpers beside the people the system keeps leaving out. A helper's profile must say what training, certification, or license that person actually has. A system nobody can take away, because I've watched what happens when they can.

If this page changed what you do with one appointment, it worked. If you want the deeper argument — the endoscopy, the raw data, the economics, the fellowship — it's in the essay: Health You Own. And the system we're building is at KISOP.

Provenance

I held this page to its own standard: claims, with sources and dates. Accessed July 6, 2026; the ChatGPT, Copilot, Verily Health, Amazon, Apple, Anthropic policy, MedGemma, and NVIDIA items re-checked September 7, 2026; the Muse section added September 8, 2026, the day of its launch.