07/Shared memory
GBrain
The Pantheon team's shared knowledge layer — identities, roles, decisions, playbooks, and what the fleet has learned. One brain, twenty readers, a single writer.
What it is
GBrain (garrytan/gbrainv0.57.0.0) is the long-term memory shared by the whole Pantheon team. Agent identities, standing facts about the box and the fleet, comms playbooks, and the lessons each agent distills from finished cards — all queryable as a knowledge graph instead of living in any one agent's head.
It is notthe message bus. Chat transport — rooms, queues, presence — lives in the Fleet Room's SQLite database. GBrain is what the team knows; the room is how it talks.
How data is saved
- PGLite, embedded. A real Postgres running inside the process — no Docker, no separate database server. Lives at
/home/hermes/gbrain-data. - Keyword search. No embedding provider, no API key, nothing to bill — recall is keyword matching over facts.
- Served over MCP.
http://127.0.0.1:3131/mcpvia thegbrain-servesystemd unit. Agents read through a helper — the CLI can't open the store while the server holds its lock. - Credentials expire hourly and refresh automatically — reads keep working with no cron and no restarts.
Write policy — Shrikanth's order
GBrain is read-only for every agent except Kratos Muse — ordered twice, 2026-09-24. All twenty agents recall their learnings page at the start of a card and never write. At card end each agent records 1–3 dated lessons in its own local MEMORY.md; Kratos Muse curates what actually lands in the shared brain. That's why there are 20 reader credentials and exactly one writer.
08/Inventory
What's inside
Live page-by-page fact counts, read from the brain itself — not hard-coded.
09/Credentials
Reader clients
One MCP credential per agent, all read-scoped. Values never leave the box.