Gstack users can optionally enable HydraDB through their existing GBrain, choose fast or thinking retrieval, and switch back to normal GBrain. This fixes the short startup timeout and text-only parsers that otherwise prevent cloud memory from reaching agents.

Changes

• Add /setup-gbrain --hydradb and --mode thinking, with separate cloud-copy consent and connection verification. GBrain handles credentials and backfill.
• Save a semantic-query budget of 10 seconds for fast or 30 seconds for thinking. Normal local setup keeps its existing default.
• Preserve readable Markdown in agent context and support structured search results, including cloud-only memory without a fabricated local page reference.
• Keep normal GBrain as the default. cloud primary off restores local retrieval; cloud disconnect stops the integration and preserves local data.

Uses the optional GBrain backend proposed in garrytan/gbrain#6695. Setup checks that capability before connecting. Gstack adds no HydraDB SDK or credential store.

Validation

• 213 focused tests passed across setup, context loading, memory adapters, search parsing, resolvers and skill parity.
• Eight native setup smoke checks passed without credentials, exercising capability detection, opt-in consent, fast/thinking query budgets, the actual GBrain CLI and Gstack context loader, and return to local defaults.
• Product typecheck and generation for all skill hosts passed; generated tracked files remain unchanged.
• Earlier live integration verified fast/thinking Markdown plus structured MCP output, cloud-only retrieval through the real context loader, updates, local switching and cloud cleanup.

Related task memory

Companion task-memory flow: #3298 adds automatic capture and recall of verified task procedures through native hooks. It is separately opt-in and works alongside the GBrain retrieval configured here.