How Context Works

Every time you send a message, Atlas assembles a “context package” before calling the AI. This context is what allows Atlas to give you relevant, personalized responses rather than generic answers. Here’s what goes into it.

What’s Included in Every Message

When you send a message, Atlas builds context from several sources in this order:

1. Your persona file — who Atlas is and how it should behave
2. Your user profile snapshot — a short summary of who you are (name, role, preferences, recent focus areas)
3. Relevant vault notes — notes from your vault that match your message, found via intelligent search
4. Recent conversation history — the back-and-forth from earlier in this session
5. Available tools — the list of things Atlas can do
6. Behavioral rules — any rules you’ve configured about how Atlas should act

The vault search (step 3) runs automatically every time you send a message. You don’t need to paste in notes or manually attach files — Atlas finds the relevant ones on its own.

Lean Context Mode (Default)

By default, Atlas uses lean context mode. Instead of cramming all your memory files into every message, it sends a minimal system prompt of around 1,500 tokens. When Atlas needs specific details — like the full contents of your facts file or a specific note — it uses tools to fetch them on demand.

This has two big advantages:

• More room for conversation — less of the context window is taken up by background information
• Faster responses — smaller prompts process more quickly

Tip

Lean context mode is on by default and works well for most people. You can turn it off in Settings > AI > Lean Context Mode if you prefer Atlas to always have all your memory loaded upfront, but this will use more of the context window on every message.

The Context Budget

Atlas reserves 60% of the AI’s context window for conversation history by default. The rest is used for the system prompt, tools, and RAG results.

As a conversation grows longer, older messages take up more of that budget. When history approaches the limit, Atlas automatically handles it:

1. Important facts are extracted first — before summarizing, Atlas silently scans the conversation and saves anything important to your long-term memory. Nothing meaningful gets lost.
2. Older messages are summarized — the early parts of the conversation are condensed into a short summary that takes far less space.
3. The conversation continues — from your perspective, the chat just keeps going.

You can adjust the context budget percentage in Settings > AI > Context Budget if you want more or less history available at any time.

How Vault Notes Become Context (RAG)

When you send a message, Atlas runs a hybrid search across your vault using both semantic meaning and keyword matching. The most relevant notes are included at the top of the system prompt as background reading for the AI.

This means:

• Atlas might reference a note you haven’t opened in weeks if it’s relevant to your question
• You don’t need to link notes or tell Atlas where to look — it searches automatically
• The more notes you have in your vault, the richer the context Atlas can pull in

Note

Only notes from your vault are used as context. Atlas never sends your notes to third parties — they go directly to the AI model via the Atlas cloud backend, and only the notes relevant to your current message are included.

Custom Instructions

You can add instructions that are always included in every message Atlas receives. This is useful for things like:

• Preferred response style (“Keep answers concise and use bullet points”)
• Topics to always remember (“I’m a software engineer working on iOS apps”)
• Behavior preferences (“Always ask before creating new notes”)

To set custom instructions, go to Settings > AI > Custom Instructions and type anything you want Atlas to always know.

---

Next: Sessions