AI & Chat Settings
These settings control how Atlas communicates with the AI model. All AI chat features require an active Talk plan or higher.
Context Window Override
Section titled “Context Window Override”By default, Atlas automatically detects the context window size for the model it’s using. You can override this value (in tokens) if you know the specific limit for a custom model or want to set a ceiling.
Leave this field empty to use the automatic default — that’s the right choice for most users.
Temperature
Section titled “Temperature”Range: 0.0 – 1.0 | Default: 0.7
Controls how predictable or creative the AI’s responses are.
- Lower values (0.0 – 0.4): More focused, consistent, and deterministic. Good for factual tasks, summaries, or structured output.
- Higher values (0.7 – 1.0): More varied, creative, and willing to explore ideas. Good for brainstorming, writing, and conversation.
The default of 0.7 balances coherence with natural-feeling responses.
Max Tokens
Section titled “Max Tokens”Default: 4096
The maximum number of tokens Atlas can use in a single response. A token is roughly 4 characters (about ¾ of a word).
- Increasing this allows longer responses for complex tasks
- Decreasing it keeps responses shorter and faster
- Very long responses (8K+ tokens) may take noticeably longer to generate
Streaming
Section titled “Streaming”Default: On
When streaming is on, you see Atlas’s response appear word by word as it generates — like watching someone type. When it’s off, Atlas waits until the full response is ready before showing anything.
Streaming is almost always the better experience. You might want to turn it off if you’re on an unstable connection and prefer to wait for a complete response.
Lean Context Mode
Section titled “Lean Context Mode”Default: On (recommended)
Lean context mode dramatically reduces the size of the system prompt sent with each message — from around 53,000 tokens down to about 1,500 tokens.
In lean mode, Atlas sends a minimal prompt and uses tools to fetch memory and notes on demand. This leaves much more room in the context window for your actual conversation.
When lean context mode is off, Atlas injects all your memory files (persona, user profile, facts, recent sessions) directly into every system prompt. This can be useful if you want Atlas to always have everything loaded without needing to fetch it, but it uses significantly more context on every message.
Context Budget
Section titled “Context Budget”Default: 60%
This controls how much of the AI’s total context window is reserved for conversation history. The remainder is used for the system prompt, tools, and relevant notes from your vault.
- Higher percentage: More conversation history fits, but less room for system prompt and notes
- Lower percentage: More room for context and notes, but older messages get dropped sooner
If you have a very large vault and want more notes included as context, try lowering this to 50%. If you want longer conversation history, raise it to 70–80%.
Custom Instructions
Section titled “Custom Instructions”Text you enter here is added to every system prompt Atlas sends. Use this for persistent preferences that you always want Atlas to follow, regardless of what you’re discussing.
Examples:
- “Always respond in Spanish.”
- “Keep answers concise. Use bullet points when listing more than three items.”
- “Never create or modify notes without asking for confirmation first.”
- “I’m a software engineer. When I ask about code, default to Python unless I specify otherwise.”
Custom instructions work alongside your persona and memory files — they’re the best place for behavior preferences that don’t fit naturally into a note.
For more detail on how context is assembled and managed, see How Context Works.