Recall

Recall maps are memory maps for agents. A map is a set of cards that an agent walks one at a time, starting at the entry card. Each card holds some knowledge and options: signposts to the next card, each with a “use when” line. A card marked Prompt is a reusable procedure that agents find by meaning.

The left side lists your maps in folders. New map creates one with its entry card. Open a map to use three views:

• Cards: edit a card’s title, body, “use when” line and options. Drag to reorder.
• Graph: the map drawn as cards and lines.
• Revisions: every change, by you or an agent. Restore puts back what a change replaced.

A save that would break the map (a card too long, an option that points nowhere) is refused with a message saying what to fix. If someone else changed the map since you opened it, the map reloads instead of overwriting them.

Three things only you can do: Publish a map so agents can find it, confirm a card as a prompt, and delete a map.

Assistant

• “Which Recall maps do we have?”
• “Open the deploy map and summarise the first card.”
• “Add a card about the backup procedure to the operations map.”

Agents read maps with recall_index, recall_open, recall_go and recall_match. Writing cards needs the Recall authoring group, which no agent has by default. A map an agent starts stays a draft until you publish it.

Technical

A map is a recall item in the tree; its cards are rows in recall_nodes, so an agent reads one card in one query. Limits: 6,000 characters per card body and 100 cards per map. Prompts are embedded for recall_match, which returns only confirmed prompts on published maps above a score floor. The last 50 changes per map are kept for restore. Changing a map’s slug keeps the old slug working. More detail: Recall maps.