Processing disclosure

How AI processing works

Encryption protects what SOL stores. An AI request is different: the material needed for that request must be readable while a service produces a response. This page explains that boundary and names the outside services involved.

Last updated September 4, 2026

The request boundary

Storage and processing are not the same thing

  1. 01

    Your stored private records are encrypted on your device

    Dreams, journal entries, memories, and chat message bodies are sealed before they are stored. SOL does not hold the key needed to open them.

  2. 02

    An AI request must be readable while it is processed

    Your new input is sent in readable form. If relevant encrypted history is included, your device decrypts that material before it becomes part of the request.

  3. 03

    The response is encrypted again before private storage

    In the standard chat path, user and assistant message bodies are encrypted on your device before database storage. Providers never receive your recovery code or encryption key.

Outside services

Who may process part of a request

The service used depends on the feature and the provider configured for it. One request can involve more than one provider. For example, one may plan a response while another produces the final answer.

Interactive artifacts run code inside a sandboxed panel. SOL asks generated artifacts to load libraries only from jsDelivr or esm.sh, but that network limit is not currently enforced. Artifact code can contact other internet locations. Do not put private information into an interactive artifact.

Anthropic

Used for
Claude responses and selected planning, title, interpretation, summarization, image-description, and voice-support tasks.
What it receives
The text, images, conversation context, retrieved sources, and settings needed for the request you start.

OpenAI

Used for
Configured AI responses and supporting tasks such as search embeddings, query rewriting, image generation, and transcription.
What it receives
Depending on the feature: text and structured context, public research material, selected images, or audio you ask SOL to process.

xAI

Used for
Grok-based answers and thread titles when that provider is configured for the feature.
What it receives
Prompt text, relevant conversation context and source results, plus a technical conversation identifier on paths that use provider caching.

ElevenLabs

Used for
Optional real-time voice conversations and spoken responses.
What it receives
Microphone audio, live transcript and response text, voice settings, and the account, thread, and session identifiers needed to run the conversation.

Deepgram

Used for
Optional live dictation that turns speech into text.
What it receives
Microphone audio streamed from your browser when you choose live dictation, using a short-lived access token.

Tavily

Used for
External web search for research questions.
What it receives
The search terms and settings SOL sends when outside web research is used.

Jina

Used for
Reading a public web page so SOL can inspect and cite it.
What it receives
The public URL requested by the feature. Jina returns a readable version of that page to SOL.

jsDelivr

Used for
Delivering browser-side embedding software and optional libraries requested by interactive artifacts.
What it receives
The requested asset URL and ordinary network information. Private text being embedded is not intentionally included in that asset request.

esm.sh

Used for
Delivering optional browser libraries requested by an interactive artifact.
What it receives
The requested library URL and ordinary network information when an artifact loads a library from this service.

Hugging Face

Used for
Delivering the small model files used by the browser-side embedding tool.
What it receives
A normal model-file request from your browser. The selected text is processed on your device rather than sent to Hugging Face.

Anthropic retention boundary

Standard API requests are not zero-retention

Anthropic says standard commercial API inputs and outputs are automatically deleted from its backend within 30 days. It may retain them longer under a different agreement, to enforce its Usage Policy, or to comply with law.

Anthropic also says commercial API inputs and outputs are not used to train its models by default unless the customer opts in or submits feedback.

Anthropic retention policy Anthropic training policy

Know what an enabled feature may add

Saving encrypted content does not itself ask an AI provider to process it. When you use an AI feature, enabled memory or search can add relevant private records to that request. Voice, transcription, web-search, and page-reading actions use the service named above for that path.

Read the full privacy policy