Review AI Sources Before They Reach Brain Memory

1. 1Inspect the source
2. 2Review its summary
3. 3Accept useful context

A link or fluent summary is not, by itself, verified evidence.

Sources let the Brain use curated material beyond the current Canvas without treating every pasted claim as truth. The product separates raw material from approved grounding.

The review-before-attach lifecycle

1. Add a raw source — a manual note, pasted text, Canvas snapshot, or supported Google Drive link.
2. Review it — write a concise summary and discrete facts, then label confidence and freshness.
3. Approve it — the reviewed summary and facts become attached memory that can ground Brain and section chat.
4. Detach or delete it — detach stops grounding but preserves the record; delete removes it.

A raw source never grounds Brain by itself. Brain uses the approved summary and facts, not the raw source's entire original text.

Confidence and freshness

Reviewers label confidence as strong, partial, weak, or conflicting and freshness as current, recent, stale, or unknown. These labels help the team decide how much weight to place on a source; they do not independently verify it.

Conflicting material should remain visible as a conflict until a person resolves it. Newer does not always mean more authoritative, and an approved source can become stale as the operation changes.

Google Drive support

Where connected-source fetching is enabled, the Google Drive connector accepts a Drive, Docs, Sheets, or Slides link shared as Anyone with the link. It fetches supported metadata and, for Google Docs, exported plain text. The imported material is still only a raw source until it is reviewed and approved. If fetching is unavailable or a link cannot be read, use supported manual notes or pasted text instead.

This is not a personal Drive OAuth connection or live two-way sync. The app does not store a Drive login or keep the Canvas synchronized with later file changes.

Inspect model memory in What I know

Open What I know from the Brain, then use the full memory page to search recorded systems, connections, facts and people. Filter by kind or confidence and inspect the source behind an item. This view reflects the model; it is separate from the attached-source lifecycle above.

A systems-model item that can be matched to a specific record can offer Confirm and Fix. Fix opens a correction conversation. Section-derived facts and people do not offer those controls; correct their owning Canvas section. Unmatched items explain why direct confirmation is unavailable. The page is not a universal editor or delete control for every kind of memory.

Conversation memory is separate

Conversation history remembers earlier Brain and section-chat sessions. It is scoped by surface so Whole Canvas, individual sections, AI Readiness, and Agentic ROI can be revisited without presenting one continuous mixed transcript.

Conversation history can help continuity, but it does not become approved external evidence merely because it was said before. Canvas edits and attached memory remain the places to correct or curate the model.

User control

Use Sources in Brain chat to manage raw sources and attached memory. Detach material that should no longer ground answers. Delete only when its retained record is no longer needed.

For how the Brain uses approved sources, see using the Agentic Brain. For the terms used here, see the glossary, and for where sources fit in the journey, how the app works.

Next: Governance and safety.