Privacy & Local AI
How Privacy Works
River keeps your work in River's cloud, so it's there on every device and River's AI can work on it. This page says plainly where your content goes and what you control. The Privacy Policy is the full version, including the list of every service that handles your data.
Where your content lives#
- Your documents, files and chats are stored in River's cloud, on Amazon Web Services in the United States. What River remembers about your work is kept by River's memory service, Mem0.
- Files on your computer stay there unless you upload them or River's AI reads them for something you asked for.
- Sign-ins of the apps you connect are kept by River or by the service River connects through (see below).
What's encrypted#
| Data | At rest | In transit |
|---|---|---|
| Documents, chats and other items | Yes | TLS |
| Files you upload | Yes | TLS |
| Search records | Yes | TLS |
| Backups | Yes | TLS |
| Sign-ins and keys of apps you connect | Yes, and encrypted again before they're stored | TLS |
Encryption at rest protects your content if someone got hold of River's disks or backups. It doesn't hide your content from River itself: River's servers read it to search it and to let River's AI work on it, and River staff with admin access can open it when they need to, for example to answer a support request.
Sign-ins of the apps you connect#
For most apps you connect, the sign-in is kept by Composio, the service River connects through, and River keeps only a reference to it. Keys and sign-ins River keeps itself (for example an API key you paste in) are encrypted on River's servers before they're stored, on top of the database's own encryption. River uses one encryption key for these, kept apart from the database.
When River's AI works#
- Your request and the content it needs go to the AI service that runs the model you picked. That includes what you wrote, files and photos you added, the items River works on, pages River's browser reads, and pictures of the windows River works in.
- Some jobs use other services whichever model you pick: the short first line that tells you River has started, the labels for the steps River takes, each chat's title, the shortening of the older part of a long chat, helpers River starts for part of a job, and one more try on another service when a model declines a request. Two more run on their own models instead of the one you picked: Quick ask, the box that opens where you point, and the greeting and suggestions that open a new chat. Quick ask receives your request, what's on your screen (your name and team, the titles and a short summary of your open windows and desktops, the text of the item you're in, what you selected or pointed at, any area you marked) and, when you ask about them, your calendar or tasks. To answer you, it can also read your documents and files, email, messages, calendar, tasks, connected apps and the web, and what it reads goes to the service that runs it. The Privacy Policy names the services. Titles and the shortening of long chats happen in Incognito chats too. A chat on a local model, or on an AI service you connected yourself, is the exception for the first six: it gets no first line or step labels from an AI model, River names it without one and keeps it in full, its helpers run on the same model, and no other model gets a try. Quick ask still runs on its own models, and the greeting doesn't read such a chat. See Local models.
- River doesn't train AI models on your content. Anthropic and OpenAI don't train on what River sends them either, and may keep requests for up to 30 days to check for abuse. Not every AI service River offers makes the same promise; the Privacy Policy says which.
- Copies of your conversations with River's AI go to two services we use to find and fix the AI's mistakes (Mirrors and Raindrop). Incognito chats and chats on a local model are left out.
- AI can be tricked. A web page, an email or a file can hide instructions meant for an AI. River's AI is told to treat what it reads as information, not instructions, and to ask you before it sends, posts or buys anything. A request you already spelled out ("email this to Sam") counts as your go-ahead. No AI is immune, so read what River is about to send.
Local models#
If you pick a local model (Ollama), the model that answers runs on your own computer, and River's own AI models write nothing for that chat:
- no first line and no step labels from an AI model: River shows its built-in wording for each step;
- its title: River names the chat from the first words of your first message, without an AI model;
- its history: River keeps a long chat in full instead of shortening it;
- its helpers: the helpers River starts for part of a job run on the same local model;
- a request the local model declines or can't answer: the request ends there, River tells you, and no other model takes over;
- its content stays out of quality monitoring, as in an Incognito chat;
- the greeting that opens a new chat doesn't read it.
A local chat is not sealed off, though. These still leave your computer:
- River's memory, whichever model you pick: if your plan includes it, each request searches what River remembers about your work, and what you wrote in a request you made yourself is then sent to River's memory service (Mem0), which keeps what it learns from it. River's replies aren't sent. Nothing from an Incognito chat is saved to or searched in what River remembers; a chat with a Venice model isn't saved, though its requests still search it. You can see and remove what River remembers in Brain;
- the tools River uses for you: when River searches the web, makes a picture, speaks, uses a connected app, or works in River's browser or cloud computer, that tool's service receives what it needs for the job;
- Quick ask, the box that opens where you point: your request, what's on your screen (your name and team, the titles and a short summary of your open windows and desktops, the text of the item you're in, what you selected or pointed at, any area you marked) and, when you ask about them, your calendar or tasks go to OpenAI and Google models, never to the local one. When your question needs it, the Google model also reads your documents and files, email, messages, calendar, tasks, connected apps and the web, and what it reads goes to Google too;
- the greeting and suggestions that open a new chat, which a Google model writes from your recent chat titles (Incognito and local chats left out), the names of your recent files, your recent activity, the apps you've connected, your profile, how you first signed up, and the names of your team and desktop;
So a local model keeps most of a chat on your computer, not all of it. The Privacy Policy names the services. The chat is still saved in River's cloud like any other chat, and the request passes through River's servers on its way to your computer. For work that must never leave a building, talk to us about River on your own hardware.
On your computer#
With River on your computer:
- River's AI can read and edit any file your computer allows, for work you ask for. What it reads goes to the model you picked, like the rest of your request.
- Pictures of the windows River works in go to that model too.
- River on your computer and River on iPhone record no sessions. On the web, session recordings keep only the layout, clicks and page changes; your content and everything you type are hidden before anything is sent. The web addresses of links and pictures on the page are recorded, with sign-in codes and keys removed.
Exporting and deleting#
- Export a document, sheet or deck as PDF, Word, Excel, CSV or PowerPoint, or download a whole folder. There isn't a one-click export of everything yet; email us if you need one.
- Delete your account in Settings › Account. Your account closes at once and is deleted 30 days later; sign in during those 30 days to keep it. Some copies last longer, such as backups and copies held by services River uses. The Privacy Policy lists them and how to have them removed.
What we never do#
- Train AI models on your content
- Sell your data
Tips#
- Use an Incognito chat for a conversation you don't want kept in your history or sent to quality monitoring.
- For sensitive work, pick a model from Anthropic or OpenAI.
- Check Your public pages in Settings now and then to see what you've published.