The agent's avatar — shown in chat messages and the presence indicator when the agent is writing.

The name displayed in the chat. This is who the agent "becomes."

The agent's identity layer — completely separate from the operational rules below. The agent will think and speak as this person without ever referencing tools or workflows.

Short keywords that define the feel. Shapes both the agent's output and inner monologue.

Off — Haiku inner monologue only. Set 1024+ to enable Opus extended thinking.

Memory is automatically saved after each run using Haiku (very cheap). The agent reads this at the start of every session so it doesn't repeat questions or forget preferences.

How it works: After each run, Haiku extracts key learnings and appends them. When memory exceeds ~8KB, Haiku compresses it — merging duplicates and keeping only the most useful items.

Firewalled from the persona layer. These are the agent's "hidden" workflow instructions — read files first, use tools, formatting rules. The persona follows these silently but never talks about them.