Build and maintain the knowledge base

The knowledge base gives agents the business context that a dataset alone cannot express. This guide explains the saved knowledge surfaces and the review inbox.

Choose the right knowledge type

Open Context and use the surface that matches the information:

Put numbers that must be computed in governed metrics, not in prose. Use knowledge for meaning, instructions, caveats, and durable facts.

Capture knowledge from work

Knowledge may be added directly or proposed from an agent conversation. Write one clear idea per entry, name its subject, and link it to a metric or dataset when that relationship matters. Avoid copying the same rule into several stores: conflicting versions make agent behavior unpredictable.

Workspace knowledge is shared context. Personal pinned facts apply only to the person who saved them. Agent-specific facts apply to that agent. More specific knowledge takes precedence, but it should not contradict the workspace's governed definitions.

Review the knowledge inbox

Open Knowledge inbox (/knowledge-inbox). The inbox collects proposals, feedback, and activity that require a decision. Open an item to inspect its source and proposed wording, then approve, edit, reject, or otherwise resolve it.

Approving a proposal is what promotes it into governed knowledge. A suggestion inside a chat is not authoritative until it passes this review.

Knowledge packs group related proposals into a review queue. Review the items; do not assume the pack itself is already active agent context.

Keep knowledge healthy

Use the knowledge health view to look for:

Resolve conflicts instead of relying on entry order. Update or remove stale items, then confirm that important knowledge is available to the agents that need it.

Related guides