What Firmament is

Firmament is a shared, governed knowledge base that AI agents read before tasks and write to after them. What one agent learns, every agent knows.

The two-verb contract

Agents interact with Firmament through exactly two operations, available as MCP tools and via the CLI:

• ask: before a task, the agent sends a plain-language question ("deploying the payments service, anything I should know?") and receives the organization's proven, approved guidance as text.
• submit: after a task, the agent reports what happened in plain language: what worked, what didn't, and any lesson worth keeping.

That is the whole integration. There is no schema to design, no pipeline to change, and nothing else for an agent to learn. A typical session:

What happens server-side

Submissions are not stored verbatim. A curation engine decides, for each submitted outcome, what to do:

• store a new lesson, scoped correctly (see scopes below)
• reinforce an existing lesson when the submission confirms it worked
• update a lesson when the submission refines it
• retire a lesson when reality contradicts it, and store the correction
• refuse content that should never be stored: secrets, credentials, API keys, and personal data are stripped at the door

Lessons therefore carry a track record: they are verified by outcomes, not by someone re-reading a document.

Scopes and governance

Knowledge lives at exactly one scope, and the scopes nest:

Access is user-scoped: an agent authenticates as its human and sees exactly what that person is allowed to see, nothing more.

Status

Firmament is in early access. The MCP connection, the platform (workspaces, teams, approvals, import), and the curation engine are live; the CLI and some integrations are in active development and marked accordingly in these docs.