# GBrain session detail — 2026-08-06

## What happened

Rob pasted a third-party doc about "GBrain + Obsidian + Hermes on VPS" and asked "which option?" The doc was AI-slop with broken commands, but the core idea was sound. I vetted it against the real environment, chose Option A (self-contained VPS stack), and installed GBrain as a semantic retrieval layer over the existing Syncthing-mirrored vault.

## Key decisions

- **Option A over B**: B (tunnel from laptop) dies when Connie sleeps. Rob's whole architecture is VPS-first.
- **Conservative search mode**: GBrain's default `tokenmax` mode is 25x more expensive. I caught this from the docs and set `conservative` explicitly.
- **No overnight consolidation yet**: `gbrain dream` spends LLM tokens. Deferred until retrieval quality proves itself.
- **Freshness cron every 30 min**: stops gateway, imports + embeds, restarts. ~5s downtime. Rob approved.

## The bridge gap discovered

Rob expected the Claude-Hermes MCP bridge to give me live visibility into his Cowork projects. Reality: it's async write-only. Claude must actively write notes to `Claude Activity/`. I created vault folders for his 6 unbridged projects; he'll attach them in Cowork and ask Claude to write summaries.

## User style notes

- Rob pastes URLs expecting me to read them. When they hit login walls (claude.ai/space/), he doesn't push back — he just pastes the next one. I should recognize this pattern faster and ask for the info directly instead of trying 3 URLs.
- Rob counts his projects carefully. I miscounted twice (said 9, then 10, then 9 again). He corrected me each time. Be more careful with enumeration.
- When Rob says "setup both," he means both options I presented, not one. Don't ask for clarification on "both."