Version: Next

Brain Runtime Topology & Deploy Model

Authority: 00_MISSION_CONTROL/runbooks/brain-services/RUNBOOK_Brain_Service_Topology.md.

Service topology with local agents on MCP 8788, relational authority, Qdrant semantic projection, and GRIFF-Map hourly and realtime maintenance. The diagram's HTTP-8787 lane is historical: that listener was retired 2026-07-16.

:::note Single-port model since 2026-07-16 The separate HTTP :8787 listener shown in the older diagram is retired (single-port directive, BRAIN_PORT_ROUTE_CONTRACT.json). Health and status are served same-process on :8788 (/health, plus the MCP initialize handshake as the deeper probe). Port 8787 is now owned by the Agent OS Framework, not Brain. :::

Plane	Live owner
Agent recall	`GRIFF-Brain-MCP`, loopback `8788/mcp`
Health/status	same process, `8788/health` (since `1.2.0.dev18`)
Runtime code	`02_MEMORY_BRAIN/runtime/griffai-memory`
Relational authority	`inventory.active.db` with WAL, FTS and graph
Semantic projection	`embeddings.db` and Qdrant `griff_memory_v1`
Structure projection	GRIFF-Map realtime delta + hourly full metadata FTS

The Brain MCP process runs High priority. CUDA is used for embeddings, semantic-edge maintenance and cross-encoder reranking. Qdrant is green and CPU-backed. Model residency can make working-set and dedicated VRAM usage look high while services are idle; judge availability by /health, /ready, lane diagnostics and the GPU/device receipts, not Task Manager utilization alone.

The final verified gates are: router matrix 11/11, full suite 1,570 passed with 11 skips, A+ 28/28, clean install 8/8 and platform boundary fail count zero. Counts such as memory items are live observations and must not be frozen as health requirements.