Privacy
Privacy policy 1.9.5: encrypted memory protects storage only
Our earlier wording overstated the encrypted memory. It protects stored entries, but entries in use go to the AI model in plain text. Version 1.9.5 says so.
The encrypted memory protects stored entries only: entries in use are sent to the AI model in plain text, so it is not end-to-end. Nothing to accept again.
We overstated the protection of the optional encrypted memory. Earlier versions of the privacy policy said the feature was not offered, and the settings and the documentation described it as zero-knowledge and end-to-end. That was wrong or misleading, and we have corrected it.
What is true: your browser encrypts each entry with AES-256-GCM before upload. The server stores only ciphertext and keys wrapped by your passphrase and your recovery code. A copy of the database or a backup does not show your entries.
What we left out: while the memory is unlocked, your browser decrypts matching entries (about 1,500 characters) and sends them in plain text with every message of a chat that is not anonymous. They pass through Cloudflare and our server and go to the model provider with your prompt, by default Featherless. This happens whether or not memory is switched on for that chat. Locking the memory stops it. The suggestion button sends the last 30 messages of your latest chat to a model. We also see how many entries you have, when they change and roughly how large they are. So this is not end-to-end encryption.
Version 1.9.5 also describes the Android phone jobs. With your consent for each job, screenshots of an allowed app and texts read through the accessibility service go to the model. Screenshots are not stored. Sharing ends with the job, pause, stop or takeover.
Nothing in the behaviour of the feature changed with this update, only the texts. You do not need to accept anything again. The full change log is on the legal page.