retrieval-reflex

Encode a trigger policy that retrieves on demand when an entity becomes salient. Open the entity's page with get_page before asserting details, and pull graph neighbors only when relationship context is needed.

Retrieve when an entity is the subject of the message, a brain-page pointer appeared this turn, an unrecognized notable name appears, or you are about to assert a non-trivial detail. Skip trivial passing mentions and anything already loaded in context.

It uses get_page to open full pages, query for search, and traverse_graph plus get_backlinks for relationship context. Over the CLI, traverse_graph and get_backlinks map to gbrain graph and gbrain backlinks.

No, it is read-only. The frontmatter sets mutating to false and writes_pages to false, so it only retrieves pages, pointers, and graph neighbors without altering any stored content.

Escalate only as far as the task needs: stop at pointer metadata for identity checks, open the full page when details matter, and pull graph or backlinks only for relationship questions. Resolve only the names the current task needs, then drop them.