Platform Overview
GRIFF AI is a multi-agent platform centered on a persistent Brain / Memory layer, an agent fleet, and a desktop control plane — built around the principles of audit, replay, and governance.
Major components
| Layer | What it is | Key surfaces |
|---|---|---|
| Brain / Memory | Persistent memory + retrieval over a single SQLite store (inventory.active.db), served read/write over HTTP and MCP. | HTTP :8787, MCP :8788, exposed via memory.griff.run (cloudflared). See Brain runtime topology. |
| Agent fleet | Claude + Codex chair agents coordinating via A2A / TeamChat. | A2A, cowork, teamchat |
| Desktop Control Plane | Electron + React operator UI. | packages/desktop-control-plane |
| Web Platform | Next.js web surface. | packages/web-platform |
| Service-ops | Reliable scheduled-task service management. | Reliable tool |
| Governance | Gates, SBOM, receipts, board. | Gate, SBOM |
Storage tiers
HOT (C: NVMe) -> WARM (B: SATA SSD) -> COLD (E: HDD). New writes start at the right tier; the platform tree lives on WARM. Never write to retired A:/D:. The Brain DB is read via HTTP/MCP only (never the file directly; SMB/sidecar corruption risk).
Operating principles
- Brain-first: consult Brain/Memory before rediscovery.
- Verify-first: check live state before claiming a fix.
- Document-as-we-build: the documentation + SBOM gate makes docs part of done.
- Receipts: every meaningful action leaves an auditable receipt (paths/hashes/exit codes), consistent with replay.