Skip to content
The Librarian

First run

Once your server is running and a harness is connected, there is nothing else to turn on. The Librarian works quietly in the background of your normal sessions. This page walks through what happens, so you know it is working.

Your agent gets a briefing

Section titled “Your agent gets a briefing”

At the start of each session, your AI tool loads a short document called the primer from your server. The primer teaches the agent the handful of things it needs to know: to recall relevant memories before answering, to remember durable facts, how to package up a handoff, and how to go private. You do not have to prompt any of this — it is delivered automatically when the agent connects.

It remembers as you work

Section titled “It remembers as you work”

You can save things to memory in two ways:

  • Just say so. Tell the agent “remember that we deploy on Fridays” and it calls the remember action for you. You can also ask “what do we know about X?” to make it recall.
  • Automatically. On supported tools, The Librarian quietly watches the conversation and extracts durable lessons on its own — with the agent making no memory calls at all. The capture hook ships enabled (you can turn it off per machine), but nothing is actually filed until you switch on the curator’s Intake in the dashboard (Settings → Curator): it is off by default, and until it is on the server files nothing it captures. Capture never sends anything while you are in private mode.

Saved memories do not appear in your collection instantly. Each one goes into the curator’s intake queue first, where it is de-duplicated, merged with what you already know, and filed — usually within moments.

You stay in control through the dashboard

Section titled “You stay in control through the dashboard”

Open the dashboard in a browser (the address server up printed, usually http://your-host:3042). This is your cockpit. Some changes the curator makes are applied automatically; the riskier ones — anything that would archive or restructure a memory — are held back as proposals for you to approve or reject. Reviewing that queue is the main everyday task, and it is walked through in Reviewing & accepting proposals.

A five-minute end-to-end check

Section titled “A five-minute end-to-end check”

First, switch on Intake under Settings → Curator if you have not already — it ships off, and until it is on the curator never processes your saves and automatic capture files nothing.

To prove the whole loop works:

  1. In your connected tool, say: “Remember that our staging database resets every night at 2am.”
  2. Open the dashboard and go to Memories. Within a moment you should see the new memory appear (it may arrive via the Proposals queue first if the curator wants your sign-off).
  3. Start a fresh session in the same or a different connected tool and ask: “When does staging reset?” The agent should recall the fact you just saved — even in a different tool. That cross-tool recall is the whole point.

If something looks wrong

Section titled “If something looks wrong”

The Librarian is built to fail quietly: if the server is unreachable, your agent is told to carry on without memory rather than block your work. So a missing recall usually means the server is down or the token is wrong, never that your turn is stuck.

  • The seven tools do not appear in your harness → re-check the MCP URL and agent token; see your tool’s Connect page.
  • Memories never reach the dashboard → confirm the curator’s intake is enabled under Settings → Curator.

Next steps

Section titled “Next steps”