Approve, Idle, or Archive Agent Knowledge

Your agent can record things it figures out so it doesn’t have to solve them twice. But nothing recorded guides future work until you approve it. This guide covers reviewing, approving, pausing, and retiring that knowledge.

Knowledge is reusable know-how recorded on your site so the agent starts from the solution next time instead of working it out again (see The Notes System). It can come about a few ways: the agent records something it worked out, you direct it to write down a specific instruction, or — most often — the two of you solve something together and you say “now that you’ve worked that out, save it as knowledge.” However it’s created, the same rule applies to whether it takes effect.

There’s a deliberate gate: knowledge doesn’t take effect until you approve it. The agent proposes; you decide what becomes lasting guidance. This is the single most important thing to understand about agent knowledge, and unlike most of Maxi, it’s enforced in the code itself — the agent genuinely cannot approve its own knowledge, no matter what.

The lifecycle

A piece of knowledge moves through a few states:

• Proposed — written down, but pending your review and not guiding anything yet.
• Active — you’ve approved it. From now on the agent draws on it in future sessions.
• Idle — paused. The agent stops using it, but it’s kept so you can switch it back on later.
• Archived — permanently retired, but preserved for the record.

Only you can move knowledge between these states. The agent can create it (as proposed) and can read it, but every approval, pause, or retirement is your call — confirmed by you in conversation.

Where knowledge comes from

You don’t have to wait for the agent to volunteer it. The most useful knowledge usually comes out of ordinary work: you and the agent tackle something tricky — a custom field, an unusual product setup, a multi-step routine — get it right together, and then you have the agent capture the result:

“Good, that’s exactly how it should work. Save that as knowledge so you’ll do it this way next time.”

The agent writes it up; it’s recorded as proposed, waiting for your approval like any other piece. You can also simply dictate a piece of knowledge outright if you already know what you want recorded. Either way, the approval step below is the same.

Reviewing what’s pending

Ask your agent to show you knowledge that’s pending approval:

“Show me any agent knowledge that’s pending approval.”

The agent lists what’s awaiting review, with the title and substance of each. Read it the way you’d review a note a new team member wrote about how your site works — because that’s effectively what it is. Is it correct? Is it how you actually want things done? A wrong or half-right piece of knowledge, once active, would steer future work in the wrong direction, so this review is worth a moment’s attention.

Approving it

When a piece of knowledge is right, tell the agent to activate it, and confirm:

“That one’s correct — make it active.”

From that point it’s part of what the agent reliably knows, and it’ll be drawn on in future conversations. As with operator notes, freshly approved knowledge applies going forward — the agent will be forced to refresh its knowledge and it will pick it up on your next request – no session restart required.

When knowledge is wrong or outdated

You have two ways to take a piece of knowledge out of play, and which you choose depends on whether you might want it back:

• Idle it — if it’s temporarily not relevant, or you’re not sure. The agent stops using it immediately, and you can reactivate it later. This is the reversible option.
• Archive it — if it’s simply wrong or permanently obsolete. It’s retired for good but kept in the record.

“That product-setup knowledge is out of date now — set it to idle.” (or “archive it”)

Either way, the agent stops relying on it right away. Note that knowledge can’t be sent back to “proposed” for re-editing — if a piece needs reworking, the cleaner path is to archive the old one and have the agent capture a fresh, corrected version for you to approve.

Why the approval gate matters

It would be convenient to let the agent just teach itself — but that’s exactly the risk. Knowledge that goes active shapes how the agent works on your site indefinitely, so a single wrong assumption could quietly propagate. The approval step keeps a human judgment between “the agent thinks this is right” and “the agent treats this as fact.” You stay the authority on how your own site works.

In summary: knowledge can be recorded by the agent, dictated by you, or captured together after solving something — but only you can approve it, and that’s enforced in code, not just convention. Ask to review what’s pending, approve the pieces that are correct (they guide future sessions), and idle or archive anything that’s wrong or outdated.