Privacy

Aria processes dictation and speech generation locally by default. Local processing does not mean nothing is saved: history is on by default, and text can also reach your clipboard, other apps, and files.

Dictations

History explains saving and retention controls. Turning saving off affects future dictations, not existing entries. Retained original audio can include background voices, even when a voice filter excludes them from the text.

Recording also uses temporary plaintext audio files. Aria deletes them when it no longer needs them. File deletion is not secure erasure and cannot remove backups or snapshots.

On a Mac, Mute while recording asks macOS to silence other apps. Aria doesn’t record, save, or send their sound.

Text cleanup with a local model

When you set a local model, Aria sends the completed text of each dictation to that server. Aria connects only to an address on your own Mac. The server decides where the text goes next. A proxy or a cloud model, such as Ollama’s cloud models, sends it off your Mac. Any program running under your account could listen on that port and read the text. What the server keeps is up to the server.

Server inference

When you use a server, your audio goes to it for recognition. Whoever runs that server can hear the audio and read the text. Pairing alone sends nothing; uploads start only when you switch recognition to the server.

Your voice profile and corrections never leave your computer. Only type my voice checks the speaker before upload when you deliver After recording stops and when you retry from history. With While speaking, audio leaves before Aria knows who spoke. The check helps recognition; it cannot promise that other voices are left out.

The server keeps no history. It deletes its working copies when each recognition ends, though crashes and backups can keep them. Encryption protects audio on the way, not from the person running the server. Pairing keys sit unencrypted in ~/.aria/inference/, readable only by your account.

To stop uploads, turn off Recognize speech on your server in Settings → Recognition, or run aria setup --local. Cancelling a dictation or revoking your device stops what comes next. Neither can take back audio the server already has.

Clipboard and apps

Copying exposes text to apps and clipboard managers that can read your clipboard. Typing inserts it into the focused app, which controls its own saving and sharing. Cancelling Aria cannot recall text already delivered.

On Linux, Aria pastes through the clipboard and the selection, so text it inserts stays on both. Aria asks clipboard managers to skip it, but they may not. The udev rule that lets the hotkey read your keyboard lets every program you run read every key and inject input too. Aria reads key presses only to tell its hotkey apart from typing, and never records what you type.

On a Mac, Aria sees which keys you press, to notice the Aria key. It never records what you type.

To place spaces, Aria reads two characters before the cursor and one after it. It skips password fields and never saves or logs them.

Chats

Chat sends your request to the assistant you choose. Claude Code and Codex may send the request and project content to cloud services. They also keep their own conversation records. Aria uses read-only settings to limit file changes, but those settings don’t prevent network access.

Local Provider sends your request to the server that provider.endpoint names. The server decides where it goes next, as with text cleanup. With history on, each request also carries the chat so far.

With history on, Aria also saves each request and its reply in history for up to 30 days. Starting fresh doesn’t delete earlier records. Clearing Aria’s history, or deleting a chat from it, doesn’t delete the assistant’s records. Deleting a chat also deletes the ID Aria uses to continue that chat with the assistant.

A chat you start from a history entry or a prompt result sends that text to the assistant with your first question. Aria sends it again whenever the chat starts a new assistant session. With history on, Aria saves one copy with the chat. With history off, Aria keeps it in memory until you delete the chat or Aria restarts. Deleting the entry or the result doesn’t delete the copy. Deleting the chat does.

Aria names a chat started on a Mac after its first question, or after the first line of the text it quotes. Clearing history doesn’t remove that name from ~/.aria, and expiry doesn’t either. Deleting the chat does.

Prompts

Prompts send your spoken words as text, along with the prompt file’s instructions, to Claude Code or Codex. A prompt you run on a history entry sends that entry’s text instead: a chat’s reply, or a transcript’s whole text. The assistant may send that text to cloud services.

With history on, Aria saves the result of a prompt you run on a history entry under that entry. Aria deletes it with the entry.

Each run starts without a project, earlier conversation, or Aria memory. Aria also disables the assistant’s tools and leaves out its saved instructions and settings. Prompt recordings are not kept in dictation history.

Memory

Memory is off by default. Turn it on in Settings → Privacy. On a Mac, Aria asks you to confirm first.

With Memory on, Aria saves the text of each request you speak in your own voice. It sends that text to Claude to find facts about you and words Aria mishears. Aria adds those facts to later chats. This includes Local Provider chats, so their text reaches Claude too. Memory needs your voice profile. Without one, Aria saves nothing.

Memory lives in ~/.aria/memory/, or the folder memory.dir names. Only your user account and administrators can read it. Aria deletes a saved request once it’s 30 days old and Aria has checked it for facts. Facts and learned substitutions stay until you delete them.

Turning Memory off keeps what’s stored. To delete it, run aria memory clear. This keeps lexicon.txt and assistant conversations. Deletion isn’t secure erasure and can’t remove backups.

Updates

The Mac app checks GitHub for a new release about once a day. The check sends the request GitHub sees from any download, such as your IP address and the app version. Turn it off in Settings → About with Automatically check for updates.

Files and folders

~/.aria/settings.toml holds engine and model choices, shortcuts, prompt names and paths, and per-app choices. It holds no recordings, transcripts, or credentials. Only your user account and administrators can read it. A paired phone can change only language, output, voice filtering, project vocabulary, and history settings.

Transcripts are separate files with no automatic expiry. History lists transcripts saved under the default folder, and deleting one there deletes its file. Choose a folder with suitable permissions and sharing settings. Delete the files when you no longer need them.

Project vocabulary scans the folder you choose. Its cache contains collected names and source locations. Disabling corrections keeps the cache; forgetting a folder removes its cached names.

Voice profiles store a numeric representation of enrolled speech locally. They are recognition aids, not a way to authenticate a user or guarantee that other voices are excluded. A profile can still identify your voice, so treat it as personal data. Backups of your home folder, such as Time Machine, include it.