Preferences

Preferences are the long-running instructions you give Dispatch about how to behave. The agent reads them at the start of every run, so you do not have to repeat yourself across conversations. Anything you want Dispatch to do consistently (tone of replies, default meeting length, who to copy on what, when to ask before sending) lives here.

The preferences surface is two files: /preferences/global.md for behavioral instructions, and /preferences/permissions.md for the policy the permission gate consults on risky writes.

What it does

• Global preferences (/preferences/global.md). Plain-prose standing instructions. Read into the system prompt every run.
• Permission policy (/preferences/permissions.md). Plain-prose rules consulted by the permission gate when a gated tool is about to run.
• Updated by editing the file. When you tell Dispatch "remember X" or "always do Y", the agent writes that instruction into /preferences/global.md directly via the file tools. Direct edits to either file are also allowed. Edits to the permissions file are gated as policy-write.

How to use it

For behavioral preferences:

• "Remember that I prefer summaries under 200 words."
• "From now on, sign off as 'A' on internal emails."
• "Default meeting length is 25 minutes, not 30."

For permission rules:

• "Always ask before sending an email to anyone outside @acme.com."
• "Don't post to #leadership without my confirmation."
• "Calendar events should default to 'tentative' unless I explicitly confirmed the time."

Asking Dispatch to "remember X" or "always do Y" is enough; the agent picks the right file.

Limits and behavior

• Plain prose, not structured rules. Both files are markdown. There is no rule language to learn.
• Permissions policy is gated. Edits to /preferences/permissions.md pass through the permission gate as policy-write, so the agent cannot quietly broaden its own authority.
• Per-organization. Preferences and permissions are scoped to the current organization.
• Conflicts. When preferences contradict each other (an earlier rule and a newer one), the agent follows the most recently written. The user can edit to resolve.

Related

• Permission gate: how the permission policy is consulted on risky writes.
• Daily notes: for time-bounded items, not standing rules.
• Cookbook: Set your guardrails: write your permission rules in one pass.

For agents

• Global preferences live at /preferences/global.md and are read into the system prompt every run.
• Permission policy lives at /preferences/permissions.md and is consulted by the permission validator on every gated tool call.
• "Remember X" / "always do Y" instructions are handled by the agent writing to /preferences/global.md directly through write_file / edit_file. The separately-named remember tool is read-only; it searches the long-term memory store, not the preferences file.
• File-write tools targeting the permissions path upgrade dynamically to policy-write risk and are gated.
• Both files are plain markdown; no structured rule language.
• Per-organization, like all memory.