Inbox & review keys

Agents write a lot. Some of it you want landing in your memory store unsupervised; some of it you want to see first. The inbox is that gate — per key, so you decide which writers you trust.

Write policy

Every API key has a write policy, chosen at creation:

• direct — writes commit immediately. For yourself and agents you trust.
• review — writes land in your inbox as pending. Pending memories are invisible everywhere that matters — timeline, search, the API, other agents — until you approve them.

Use review for new agents you're still calibrating, and for bulk ingestion like meeting-transcript pipelines, where one session can produce a dozen writes of uneven quality.

Triage

The inbox at Inbox in the app is built to be fast — approve or reject a session's worth of writes in seconds, keyboard only:

Key	Action
j / k	Move between pending memories (focused row auto-expands)
a	Approve
d	Reject
e	Edit, then ⌘↵ to approve the edited version
x	Select · shift+j/k range-selects
z	Undo — there are no confirmation dialogs

Each pending write shows which key and source_label produced it, and per-source chips let you bulk-select everything from one agent at once.

The fine print

• Approval is web-only, by design. There is no API or MCP surface for it — an agent can never approve its own writes.
• Agents see status: "pending" on a review-key save. That's normal; the agent prompt tells them not to retry.
• A pending memory doesn't age. Its strength clock starts at approval, so time spent queued costs it nothing.
• Pending writes count against the monthly allowance when created; rejecting doesn't refund.