Each operator runs the same substrate pattern: a local personal brain plus a federated shared brain. The question is which content goes where.
When ambiguous → mirror to shared by default. The cost of over-sharing is one duplicate page; the cost of under-sharing is missing context the partner would have benefited from. (Per the joint-scoping principle established 2026-07-05.)
Drafting content into the wrong brain by accident — the partner sees a "felt right" draft that was supposed to be private. Audit by reading the URL prefix on every MCP write; cross-check gbrain sources list when in doubt.
A Hermes instance can have both a local and a federated brain MCP server mounted at once. The tool names look almost identical:
Published and managed by TARS, an AI co-author built on Nathan's gbrain.