What the chat knows about you

The chat assistant isn’t talking to you cold. Every turn, Moonjar assembles a context — a block of facts about you and your library — and includes it in the prompt sent to Claude. This page is what’s in that block, and what is and isn’t shared with the AI providers Moonjar uses.

What context Moonjar sends each turn

A short context block is prepended to your latest message. It contains:

Date and time — current, in your timezone.
Location — your phone’s last known location, if you’ve granted location permission. Used by tools like weather and nearby places, and to plot map points.
Your collections — names and counts. Lets the assistant pick the right one without asking.
Recent and relevant memories — see below.
Pending notifications.
Active trip — if you have one in progress, its name, dates, and destination.

If you’ve focused the chat on a specific document, collection, or trip, that thing’s full content (or summary, for collections) is appended on top.

Memories

Memories are the persistent facts the chat reuses across conversations. Each turn, Moonjar pulls in:

Recent memories — the most-recently-saved set, capped to keep the context small.
Semantically relevant memories — Moonjar runs an embedding search on your message and pulls in memories that match.

In a shared chat, other participants’ memories are included too, with attribution — so the assistant knows what each person has and doesn’t accidentally reveal one user’s memories to another.

Custom instructions

Anything you set in custom instructions gets appended to the system prompt — it shapes tone, length, and style on every turn.

Which AI providers see what

Moonjar uses three AI providers. Each sees a different slice:

Anthropic (Claude). Sees your messages, the assembled context block, your custom instructions, and the results of any tool calls the assistant makes during a turn. Used for the chat itself and for document classification at capture time.
Cohere. Sees the text of your captures and your search queries (used to generate embeddings). Cohere is only used for embeddings — it never sees the chat.
OpenAI. Sees different things depending on which voice surface you use:
- One-shot voice input (the mic button in the chat composer) and spoken replies: OpenAI sees only the audio. Audio in, text or audio out. None of your captures, memories, or library are sent.
- Voice mode (the live conversation in Conversation Mode): OpenAI sees the same context block the text chat does — your collections (names + counts), your last five documents (titles), your most-recent memories (up to 20), your location if granted, and your custom instructions. This is needed for the assistant to answer “what’s on my plate” without a tool round-trip. If you’d rather keep that data out of OpenAI’s reach, use the text chat instead — Anthropic sees that context, OpenAI doesn’t.

You consent to these providers when you create your account. The AI consent page explains and lets you review.

What doesn’t leave Moonjar

Your password. It’s hashed, never sent to any AI provider.
Your API key, if you have one (programmatic access).
Other users’ data. A shared chat shares the participants’ conversation; it doesn’t share their full libraries.
Memories of other users, except the slice surfaced inside a shared chat as described above.
Documents that aren’t relevant to the current turn — the context block carries metadata and recall, not the entire library.

How long the assistant remembers

Within one conversation: every prior message is sent to the model on the next turn. The assistant has the whole thread.
Across conversations: only memories. Nothing else from a previous chat is automatically present in a new chat unless you explicitly reference it.

After a 5-minute idle window the prompt cache used to keep token costs down expires and is rebuilt — this affects cost, not what the assistant can see.