Safety & privacy
What happens to what you tell it.
Upstream is for the kind of conversation people don’t have anywhere else. That puts a real obligation on it, and most of that obligation is about being clear on what it won’t do.
If things are bad right now
On a crisis message, Upstream stops.
Every message is checked before it goes anywhere. If it matches a crisis pattern — suicidality, self-harm, overdose, passive ideation — the model is never called. Not called and then overridden: never called. You get a fixed block of real resources, written in advance, and the message is flagged permanently as what it was.
The check errs toward over-triggering on purpose. A safety net that occasionally catches an ordinary sentence is doing its job better than one that misses a real disclosure.
None of that makes Upstream a crisis service. It isn’t one, nobody is monitoring it, and it should never be the thing standing between someone and help.
Right now
In the US, call or text 988, any time.
If someone is in immediate physical danger, call 911 or your local emergency number.
No account needed to see that page.
Where the lines are
Three things Upstream is not.
Not therapy
No diagnosis, no treatment, no clinical vocabulary, no session notes. Upstream doesn’t use the word “client” and doesn’t keep a chart. If what you need is a therapist, it will not stand in for one — and having both is a good idea.
Not advice
It gives no medical, legal or financial advice. That boundary is written into the instructions the model runs on, not just into this page. It asks the question you’ve been steering around; the decision stays yours.
Not a crisis service
It is not staffed, not monitored, and not watching for emergencies. Nobody is on the other end. What it does instead is described below.
Your data
Kept small on purpose.
What’s stored
Your conversations, the wrap-up notes made from them, and the profile answers you gave during intake. That’s what the product needs to work, and it’s what there is.
What carries forward
Patterns, insights and open steps — not transcripts. The next conversation starts from what you worked out, not from a replay of what you said.
What you can see
All of it. Everything Upstream holds about you is visible in the app, and any piece of it can be let go without explanation.
What we don’t do
We don’t sell your data, we don’t run ads against it, and we don’t report your use to any third party — no employer, no program, no officer, no funder.
Upstream sends your messages to a language model to generate a reply — that’s how it works, and it’s named plainly rather than buried. The privacy policy covers the specifics, including sub-processors and retention.
Memory
A hypothesis, not a dossier.
The difference between a product that remembers and a product that surveils is mostly posture. Upstream is instructed to treat what it remembers as provisional, not to recite it back at you, and to let you correct it. When it has you wrong, saying so is enough — you don’t have to go and edit a profile.
That wording is in the prompt itself. It’s the difference between a friend who remembers and something considerably worse.
If something here is wrong, tell us.
That includes a reply that missed something serious, a resource number that’s out of date, or anything on this page that doesn’t match what the product actually does. The gap between the two is the thing worth knowing about.