memory-tasks — Claude Code skill: tasks that survive context

What memory-tasks does

This skill stores tasks as notes with type: Task. They sit in the Basic Memory knowledge graph, are checked against the schema/Task.md schema, and outlive context compaction. Fields such as status, assigned_to, current_step and steps go into both metadata and observations: metadata drives bm_search filters, observations drive bm_schema(command="validate").

A task is created with bm_write using note_type="Task" and a metadata block. The body carries Observations, a Steps list with checkboxes, and a Context section. Write that context as if explaining the work to someone who knows nothing about it — file paths, decisions already made, what you tried. After a session is compacted, the skill searches for active tasks with bm_search(note_types=["task"], filters={"status": "active"}) and picks up from current_step.

Before editing, read the task as JSON and pass its checksum as expected_checksum. Two sessions often touch the same task, for instance during a pre-compaction flush. A revision conflict means the other session wrote first, so re-read before writing. The skill advises one task per unit of work and closing finished ones instead of leaving them active.