AI privacy: what encrypted sync protects

An AI app can protect your stored conversations and still send a prompt to an external model to answer it. These are different parts of the same workflow. To understand privacy, follow the information from the moment you type to the moment it’s stored, synchronized, or passed to another service.

You don’t need to become a cryptographer. You need clear answers about who receives each piece of information, who holds the keys, and how you can remove access later.

Start with five separate questions

Use this checklist when comparing assistants:

1. Where does the app store information on your device?
2. How does it synchronize that information to other devices?
3. Which system receives the prompt to generate a response?
4. Which connected tools receive data during the task?
5. What happens to files, backups, and retained copies?

Ask these questions for your actual workflow. A short public question, an uploaded work document, and a tool that sends an email each send data to different places. A general privacy slogan doesn’t answer all three. When you decide, save the provider’s current explanation with the date you read it, so you can spot changes later.

Distinguish transport, storage, and sync

Encryption in transit protects data while it moves between two points, such as your phone and a server. Encryption at rest protects data while it’s stored. End-to-end encryption adds one more question: who holds the keys? If only your own devices hold them, the service in the middle can’t read the content.

For background, MDN explains how TLS protects connections, and Signal explains end-to-end encrypted messaging. These protect different things. A lock icon alone doesn’t tell you where a service can read your information.

For device sync, ask how a new device gets approved, how the keys reach it, and what you need to recover your account. A paired device still needs a screen lock and current software. Encryption during sync doesn’t stop someone with your unlocked device from reading what’s already on it.

Follow the prompt to the model

Inference is the step where the model writes its answer. If a remote provider does this, find out what that provider receives and which terms apply. Consumer subscriptions, APIs, and work accounts can have different settings. Check the exact product you use, not the promises of a different plan.

Suppose you ask for a summary of a client proposal. Before uploading the file, check whether the full document is sent, whether the app adds conversation history or memories, and whether a tool retrieves more context. Remove names or unrelated sections when they aren’t needed for the task.

A custom endpoint needs the same check. Find out who runs it and whether the app falls back to another model when it fails. A local-model setting doesn’t mean that every feature, sync request, or connected tool works offline.

Treat tools and files as separate destinations

Connecting a calendar, document service, or MCP server adds another participant to your workflow. Check what each action sends and what access you grant when you sign in. A read-only lookup needs less review than an action that publishes or emails something.

The MCP architecture documentation is useful background for understanding those connections. Our MCP setup guide turns that distinction into a small, observable connection test.

Ask separate questions about file storage. Having to sign in to open a file is access control; on its own, it isn’t end-to-end encryption. Check file encryption, previews, share links, caching, and deletion separately from how the app syncs chats. Ask the same questions about Mono’s Vault or any other file feature you use.

Make memory and deletion concrete

Saved memories, chat history, downloaded files, and provider records are different things. Deleting a message from a chat doesn’t tell you what happens to every copy made while answering it.

Before you rely on a temporary or private chat mode, find out whether it reads your existing memories, saves new ones, keeps local history, or changes how the provider handles your data. In Mono, Secret Mode chats never write to memory, and you choose whether they read existing memory at all. Then check the delete controls for your account, paired devices, connected services, and backups. If something isn’t explained, ask before you share sensitive content.

A practical monthly check is to review paired devices, remove unused integrations, inspect saved memories, and verify the recovery method you intend to use. Keep recovery information outside a device you could lose. Choose a retention period that matches the material you want to keep, rather than keeping everything by default.

Apply the checklist to Mono

Mono encrypts conversations, memory, and settings on your device before they sync, so the server can’t read them. Still, treat model processing, connected tools, and file storage as separate flows and check each one. Mono Code shows another example: encrypting traffic between your phone and desktop doesn’t change what a coding agent sends to its model provider.

Start with a low-sensitivity task and check where its data goes before you give the app more access. This works whether you use a hosted model or a custom endpoint. Mono is currently on a waitlist. Join the Mono waitlist for updates, and use the same questions on the tools you use today.