Skip to content
MantleMantle docs

Recall maps

A Recall map is a small set of cards you write for agents, so they follow your way of doing a job instead of guessing it from search results.

How a map works

  • A card holds one piece of know-how, up to 6,000 characters.
  • Each card has options: signposts to other cards, each with a “use when” line. An option can also lead to another map.
  • Every map starts at its entry card. An agent opens the map there and follows the options that fit its task, one card at a time.
  • A card can be a prompt: a reusable procedure agents find by meaning. An agent describes its task in a line, and the closest prompts come back.

Agents read maps with four tools: recall_index (list the maps), recall_open (open a map), recall_go (go to a card) and recall_match (find a prompt). Claude and other MCP clients get the same tools.

What stays yours

An agent may add and edit cards, and every change is logged under Revisions, where you can restore it. Three things only you do:

  • Publish a map. Agents see only published maps.
  • Confirm a prompt. An agent can ask for a card to be a prompt, but it never matches until you confirm it.
  • Delete a map.

If an agent changes the words of a confirmed prompt, it waits for your confirm again.

Example

You write a map called “Release a new version”:

  • The entry card says: “Releases go from main only. Pick the step you need.”
  • Option “use when the tests have not run yet” leads to a card with the test command.
  • Option “use when tests pass” leads to a card with the release steps.

The release card is a confirmed prompt. Weeks later you tell Claude Code, “ship this fix”. It runs recall_match with “release a new version”, gets your card, and follows your steps instead of inventing its own.

Where to edit them

Open Recall in the menu. New map makes a map with its entry card written. The Graph view draws the cards and their options (Screen help: Recall).

Next