Does Your Health Chat Remember the Last Question? Stateless Ask vs Memory Across a Session
Health chats have started selling memory. The app holds your last question, your last answer, the whole thread, and folds it into the next reply. It feels like a smarter assistant, and it is. It is also a feature that only exists because a service somewhere is keeping a record of you. Here is what that memory buys, what it costs, and how an answer computed fresh on your phone changes the trade.
AI-assisted, human reviewed. How we use AI
Quick answer
"Memory across a session" means the service stores your conversation, and any health numbers in it, and replays them into the next answer. That makes follow-ups feel natural, and it puts a transcript of your health questions on someone's servers. A chat that computes each answer on your own device, from your own data, and shows the samples and window behind it does not need to remember you to give you a checkable answer.
What "memory" actually means in a health chat
When a health chat says it remembers, it means stored state. Somewhere sits a conversation record: your question, the answer a model wrote, often the numbers pulled in to produce it. On the next turn that record goes back to the model alongside your new question.
That is useful. It is the difference between "how did my sleep change over the last month?" and "and how about last month?" It turns separate questions into one conversation, and rival apps have started to headline it: memory across chats, a repeatable analysis workflow, a conversation that builds.
Notice what the feature is made of, though. Memory is not intelligence. It is a store of your past questions, and it is the one part of the product that cannot be computed fresh, because its whole value is that it was kept.
Why memory is a server feature
Anything that remembers has to be written down where it survives between sessions, and for a chat that is a server. The transcript outlives the moment your screen is open, so it is persisted and handed back to a model on the next turn. There is no version of "memory across a session" where nothing is kept.
So the real question is not "is memory nice" but "where does it live, and who is it about". A memory store keyed to you is a profile of the health questions you ask: the thing you were worried about at 11pm, the number you kept checking. A trail, even if it is not a health record.
MetricBridge takes the other path. Every Ask answer is computed on the iPhone, from your Apple Health data, and rendered as a provenance card. There is no account and no sign-in, and nothing reaches a server we operate, so there is no server-side conversation to remember, and none to leak. Your health data never reaches us or a cloud AI.
What a stateless answer buys instead
The trade is that each answer stands on its own, and a good answer that stands on its own is one you can check. Every MetricBridge answer shows the days, the samples and the computation behind it, one tap on the card. You do not have to trust that the assistant remembers the right context, because the context is printed on the answer.
Here is a real card, from our own seeded 30-day series, so you can see the shape of it:
up 16.3% versus the prior 14 days
samples: 7 · window: 5 Nov to 11 Nov
identical on a re-run · sha256 f2e87621
The headline number, the direction, the window, the count of samples, and the same result when you ask again. That last line is the one memory cannot give you. A remembered answer and a recomputed answer can drift apart; a provenance card is reproducible by construction, because the phone runs the same calculation over the same samples. Asking costs nothing, so there is no meter nudging you to save an answer instead of re-deriving it.
The honest case for memory
It would be cheap to pretend memory is only a privacy problem. It is not. Memory wins on two real things:
- Follow-ups are shorter. You can ask "and last month?" instead of restating the subject and the window.
- The chat can build a picture over time. A service that remembers can blend several questions into a longer narrative, the kind you get from a coach who has been following along.
If you want a chat that behaves like a running conversation, a service with memory is a fair choice. Pick it knowingly. The cost is the record, not the feature.
The memory you cannot see
The part worth pausing on is not the memory itself, it is the visibility of it. A provenance card shows the numbers behind an answer; there is no equivalent card for a conversation store. You cannot see what was kept, how it was keyed, or how long it survives. One design keeps a record you take on faith; the other shows the working of every answer and keeps nothing.
None of this makes a remembered chat unsafe. It makes it a data path, and data paths should be named out loud. MetricBridge says where data goes when it goes anywhere: to the destination you pick, or to your own Mac when you turn on Your Mac mode. The MCP server makes zero network calls.
The one-line test. Ask the app where your last question lives. A memory-first chat has to answer "on our servers, in your account". An on-device chat answers "on your phone", and can show you the number, the window and the samples behind the answer in the same breath.
Four questions to ask any health chat
- Does it remember, or recompute? Memory means a record; recompute means each answer is derived fresh from your data.
- Where does the transcript sit? If the feature needs an account, the transcript is on the vendor's side of the line. Ask how you see it and how you delete it.
- Can you check one answer? Pick a number the chat gave you. Can it show the samples, the date range and the calculation? If not, memory is doing the work provenance should.
- What happens on a re-run? Ask the same question twice. A deterministic on-device calculation returns the same figure. A remembered reply can quietly move.
The price of remembering, one time against a meter
A memory store has to be hosted, and hosting is a bill somebody keeps paying. That is why the products built around a remembered conversation tend to be subscriptions, with a metered prompt count. MetricBridge has no server of ours in the path, so there is nothing to meter: a 7-day free trial, then a single $24.99 one-time unlock that covers the export, the MCP tools, on-device Ask with a provenance card on every answer, the Weekly Health Brief, and the Mac and Apple Watch apps.
For the design behind the card, see how Ask runs on the phone and the published methods behind every number. For the other half of the surface, the MCP server hands the same data to your own agents with no server in between.
FAQ
Does MetricBridge remember what I asked last time?
MetricBridge computes each answer on your iPhone from your own Apple Health data, then shows the samples, the date range and the calculation behind it. There is no account and no server we operate in the path, so there is no server-side transcript of your questions. Every number is reproducible from the provenance card on the answer itself. Wellness information, not medical advice.
Is a health chat with memory more accurate?
Memory changes the conversation, not the arithmetic. It lets a follow-up say "and last month?" without you restating the subject. Accuracy comes from the numbers and the window behind each answer, which is what a provenance card shows. A chat that remembers you can still hand you a figure you cannot check.
If a health chat keeps no history, do I have to repeat myself?
Somewhat, yes, and that is the honest cost. A stateless question has to name its own subject: "my HRV this week versus the two weeks before" rather than "and last month?". You trade a little convenience for answers that do not depend on a record of you existing anywhere.
Where does the health data go when a chat remembers?
A memory feature has to be stored somewhere, so the conversation, and any health numbers in it, sits on the service's servers for as long as it is remembered. That is a data path, not a badge. Ask where the transcript lives and how you delete it before deciding whether the convenience is worth the copy.
Ask your health data on the phone, and see the working
Every answer computed on your iPhone from your own Apple Health data, every number shown with its samples and window. No account, no server of ours in the path, nothing to remember.
Where it goes: the app reads Apple Health read-only and computes answers on the device. Your health data never reaches us or a cloud AI. What you get: all 194 metrics, on-device Ask with a provenance card on every answer, the Weekly Brief, the export and the MCP tools (21 tools, 19 read-only, and 23 prompts), and the Mac and Apple Watch apps. Price: $24.99 one time after a 7-day free trial, not a subscription. Wellness, not medical advice.