The memory map
An agent's page: what it knows, what it taught itself, its tasks and its tools, drawn from what is live.
Part of the crew rollout: Agents shows as Soon in the app until it reaches your account. See Hiring your crew.
Open an agent and you land on its memory map: a picture of its mind, drawn from what is live. The agent sits in the middle. Around it are what it knows, what it taught itself, the work it is given and the tools it works with.
Co Builder sits beside the map. You don't rebuild the agent on the map: you tell Co Builder what should change, and it shows the change as a card (see Changing it with Co Builder). The two small edits the map offers itself, switching a call off or on and renaming a path, go on that same card.
Reading the map
| Where | What sits there |
|---|---|
| Above | What it knows: facts, the people it works for, notes and documents. |
| Below | What it taught itself: the rules its runs wrote down, and its skills. |
| Left | Its tasks: schedules, chats, webhooks, you, work handed over by other agents, and test runs. |
| Right | Its tools, the agents it hands work to, and (once a run has searched them) its own past runs. |
| Beside it | What it runs with each message: its calls and decisions, on the left before it thinks and on the right after it answers. Its other agent steps are small dots on its ring. |
What is more present in its mind sits nearer the agent: first what is in every prompt, then what it opened lately, then what it hasn't opened.
- Hover a mark to read it and light what it's tied to.
- Click to hold its card open, with what you can do from it; click the empty map to let go. Enter does the same on a mark you tab to, and Esc lets go.
- Drag a memory to pull it.
While it runs, the agent and the task that started it light up.
The "?" beside the controls
The ? next to the map's controls explains the map for this agent:
- In this agent: how many memories, rules it wrote, skills, and files it hasn't opened lately. Pick one to light it on the map.
- Changes waiting for you: a run that may not write the agent's memory proposes the change instead. You keep or discard each one under Knowledge, or ask Co Builder about them.
- Paused: why, and Resume when you can.
- What it remembers of a conversation: by default the last 20 exchanges of each conversation. Turned off, it starts every message fresh, with its memory files and nothing said before.
- What each kind of run may do with its memory: which runs write it, which only propose a change for you to approve, and which only read it. A test run reads it and changes none of it.
- Its reflexes: four checks, one at each side of the agent. Relevance picks which skills and memories reach a run's prompt, completeness decides whether a run is done, self-grade grades how well a run did, and context prune decides which old tool results a long run keeps. Each says whether it is on for this agent.
- How a skill is kept: a skill the agent wrote itself is archived once it has failed more often than it worked over five runs or more. One not offered for 30 days goes stale, and is archived 60 days later. A skill you wrote is never retired automatically.
- Marks: what every mark on the map means.
What the agent knows about one person stays with that person: your private notes are loaded in the runs you start, a chat sender's notes in their own chats, and neither is offered to anyone else's run.
What else the map tells you
- Its steps: a call names only where it goes, for example "Reads from api.openweathermap.org", "Sends to hooks.slack.com" or "Sends its answer to api.telegram.org", with how many recent runs used it and how many failed. A decision lists its paths, where each leads and how often each was taken; a dotted line runs to where each path leads, thicker for the one taken most, and a red dot says when a path could not be checked. Hold a call's card open to Switch off or Switch on, or a decision's to Rename a path; the edit goes on Co Builder's card and nothing changes until you apply it. A switched-off call is skipped: what came in goes on unchanged. Ask Co Builder is on every card for anything bigger.
- Removed steps: a kind of step that was removed from Modus is said in the map's corner, with what it does now: nothing, or that the agent can't run until it's taken out.
- Tools that no longer exist: when the agent lists tools that were removed from Modus, the map names them in one sentence. Every run skips them without an error. Take them out puts their removal on Co Builder's card; nothing changes until you apply it.
- A change on its way: while Co Builder has a change open, the map draws it over the live agent. What it would add shows as Proposed, what it would take away as Removed if applied, and what it would change carries a dot. Apply or discard it and the map settles.
- Nothing live yet: a new agent turns every run down until it has a live version. Tell Co Builder what it should do; its first change goes live when you apply it.
Live version, Versions and Undo
The bar above the map says Live · version N: the version its runs use now. Every change you apply makes a new version.
Versions lists the recent changes and every version:
- Undo on the latest applied change takes it back.
- Pick a version to look at it as it was. An old version is only looked at; Co Builder changes nothing on it, and the map offers no Switch or Rename on it.
- Restore makes an old version live again. It is recorded as a change, so Undo takes it back too.
Pause on the same bar stops its schedules, webhooks and chats from starting it; a run you start still works. Resume starts them again.
Where an agent starts
- Hire an agent: describe the job in a sentence. The agent is made and you land here, with the job waiting in Co Builder. Build it starts the setup, which comes back as a card to apply.
- Browse templates: each card lists what the agent would do, step by step. Picking one opens it here, and Create makes the template its first live version.
- Build it with Co Builder: start from nothing and describe what it should do.