using-second-brain

State what to save and, if known, the destination. The agent searches for an existing record first, saves the smallest useful record with a stable source pointer and title, then verifies the write by readback or a durable receipt. Explicit save requests are honored without requiring proof of usefulness.

Ask the agent to continue the named project. It retrieves the objective, canonical task state, recent decisions, and blockers, then advances the nearest unfinished unblocked action within existing authorization, recording progress and a next step when durable maintenance is in scope.

No. The workflow is agent-, vendor-, and storage-agnostic. It discovers the available tools, folder layout, and conventions through a Brain Contract rather than assuming any app, database, metadata format, or automation system.

The agent uses a capability fallback matrix: it saves to a designated destination with a stable identifier, keeps storage-native receipts or version IDs, and explicitly reports that duplicate detection, readback, or discoverability was unavailable instead of claiming success.

Skip it for transient answers, scratch work, typo fixes, or small implementations whose context is already supplied. Retrieval alone is appropriate when implementation only needs prior decisions from notes; that does not authorize reorganizing the workspace.

Not from a broad cleanup request alone. The agent inventories first, proposes candidates with rationale and impact, and requires explicit authorization for deletion, moves, merges, bulk rewrites, taxonomy changes, migrations, or automation.