Operate your workspace's knowledge base: browse and edit concepts, run acquisitions, review proposals, and govern what your agents read.
The Tenant Brain is your workspace's versioned business knowledge, stored in the Open Knowledge Format (OKF). Your AI agents read what's published and propose what they learn. You operate all of it from the Brain Browser at Agents → Workflows → Tenant Brain (/knowledge/brain).
How much publishes without a human is your decision, not a fixed rule. By default a proposal waits for review — but under the autonomy settings on the Governance tab, a proposer that has earned trust (or a high-confidence, verified proposal) can publish immediately, with a window in which you can undo it. Every surface reports which of the two actually happened; none of them promises a review that your settings will not perform. See Reviewing what your team teaches for how to set and audit this.
Access is permission-gated: users need knowledge:brain:read to view and knowledge:brain:manage to publish, reject, or run acquisitions.
Want to correct an agent rather than operate the knowledge base? That is the training loop, and it needs no permissions at all — start at Training your AI agents.
| Tab | What it's for |
|---|---|
| Browse | Browse and edit every concept, organized by directory. Drafts and edits stay private until a human publishes — agents only read published revisions. |
| Acquire | Build knowledge from your own sources — crawl your website, upload documents, or scan CRM and Drive. Everything acquired lands as a cited draft in Proposals. |
| Proposals | Review what agents, MCP tools, and acquisition runs propose, and approve or reject each one. Anything held for review waits here — under higher autonomy settings some proposals publish without queueing, and the Auto-published list shows those. |
| Graph | See how concepts, agents, and CRM records connect — which agents read which knowledge, and what each agent proposes. |
| Gaps | Topics your agents are asked about but the brain can't answer well, ranked by how often they caused a failure. Acquire straight from a gap to close it. |
| Log | The full audit trail — who published, edited, rejected, or rolled back which concept, and when. |
| Sync | The embedding indexes behind the brain's advisor (search_knowledge) and the Retrieval Lab: one entry per published concept's handles (title, description, tags, aliases, headings) and one per section. Agents see a map of the brain in their instructions and read entries by path either way; published knowledge is no longer copied into your vector stores. |
| Governance | Graduated autonomy: how much trust an agent must earn before its proposals can publish without review. Off by default; every change audited. |
Every agent form (voice, chat, SMS, mail, task and document) has a Tenant Brain block on its Tools tab, under the tools list. It sets, for that agent alone: whether the brain's map rides in its instructions, lean or full map lines, the map's character budget, one read's character budget, how many entries a lookup returns, and whether a lookup that names no path gets the advisor's pick served or the shortlist returned. Each has a platform default and a per-channel recommendation (a voice agent wants a smaller map and shorter reads than a task agent). The voice guide walks through the block: Brain settings for this agent.
On the Acquire tab, each source runs on demand (Mine now) or on a schedule (Save scheduled mining):
Every run's results land as drafts in Proposals, each with a citations section and an origin stamp showing which run produced it.
The Proposals tab is the human gate. For each draft you can read the content, its citations, and its origin (which agent, call, chat session, or scan wrote it), then Publish or Reject. Rejecting an update to a live concept rolls it back to the published revision — a live concept is never archived from the queue. Retrieval-archive items (marked alpha) approve into the agents' vector store instead of the concept bundle — review those carefully.
The Gaps tab turns real conversations into a to-do list: topics agents were asked about but couldn't answer well. Each gap links straight into an acquisition, so closing a gap is one flow — spot it, mine it, review the draft, publish.
Training your agents — the other half of this loop. Your team corrects agents where they spoke, and those corrections arrive here as drafts (or as published knowledge, depending on your autonomy settings):
Also: