1 — What Brain Is
Brain is not a single service — it is a coordinated stack that gives an agent durable memory, structured knowledge, and routed model access. The central instance runs on ArcHelm and is managed by Agent Space. Each client gets their own isolated copy provisioned from a snapshot of the main Brain at delivery.
The main Brain compounds over time. Client Brains are frozen at delivery. That gap is by design — we always stay ahead, clients get a known-good baseline, and the gap widens with every release cycle.
neo4j / ArcadeDB
The knowledge graph. Entities, relationships, and observations across every domain. Multi-database per tenant on ArcadeDB (Bolt-compatible). Per-project neo4j instances on ArcHelm K8s for internal use.
atlas-memory
Three-tier memory: episodic (per-turn), semantic (consolidated facts), procedural (task recipes). Hybrid BM25 + vector retrieval with RRF fusion. ONNX embeddings — no LLM in the pipeline.
neo4j-mcp
MCP server exposing the graph to agents via
create_entities, add_observations, search_memories, and read_graph. One instance per project on an assigned port.ArcRouter
Model gateway. Routes all inference through the Brain, enforcing tier entitlements, metering, and fallback chains. Port 4001 on ArcHelm. 21 enabled descriptors across 3 rigs + Anthropic cloud.
Brain API
Tenant-scoped REST + MCP façade. Agents authenticate with a
sk-... key from ArcHub. All reads and writes pass through it. Port 8050 on the Brains droplet.ArcHub
Model artifact registry + tenant registry. Postgres metadata + MinIO blob store + HuggingFace shim. Issues API keys, stores Noema fine-tuned models, manages the BRAINLIC release channel.
2 — Network Diagram
This diagram shows the target architecture. The Brains droplet has not yet been provisioned and no client is syncing to it yet. Swift Innovation has local neo4j + ArcRouter running on ArcHelm; ADage1, Funding Forge, and Packly still need the full Brain stack (neo4j, ArcRouter, atlas-memory, neo4j-mcp) deployed to their hardware. All four need their sync jobs wired to Brains once the droplet is up.
Brains Droplet
Not yet provisioned
ArcadeDB / neo4j
Multi-tenant graph · one DB per client · Bolt-compatible
Brain API (port 8050)
Tenant-scoped MCP + REST for all clients
atlas-memory (central)
Aggregated session nodes across confirmed tenants
ArcHub
Key issuance · tenant registry · snapshot releases
Confirmed tenants
Swift InnovationT3
ArcHelm · 10.20.10.29
neo4j (local)the-grid graph · port 8000 ✓
atlas-memoryrunning on ArcHelm ✓
ArcRouter3-rig fleet · port 4001 ✓
neo4j-mcpport 8000/mcp ✓
Brains sync pending
ADage1T2
Client compute · needs deploy
neo4jnot yet deployed
atlas-memorynot yet deployed
ArcRouternot yet deployed
neo4j-mcpnot yet deployed
Stack + sync pending
Funding ForgeT2
Client compute · needs deploy
neo4jnot yet deployed
atlas-memorynot yet deployed
ArcRouternot yet deployed
neo4j-mcpnot yet deployed
Stack + sync pending
PacklyT2
Client compute · needs deploy
neo4jnot yet deployed
atlas-memorynot yet deployed
ArcRouternot yet deployed
neo4j-mcpnot yet deployed
Stack + sync pending
Pipeline — pending close
Star Seeds
Beyond-Sleep parent · Taiwan · Idea Tavern
Avalon
Cliff Smith · hair viz · T3 · BL-R8 open
Ag Express
Eric Randolph · ag-electronics · MVP Q1 2027
Aris
Prospect · tier TBD
SoulAura
Prospect · tier TBD
3 — What Runs Where
| Layer | Brains Droplet (central) | Client hardware (local) |
|---|---|---|
| Graph DB | Multi-tenant ArcadeDB — one isolated database per client. Cross-tenant queries return aggregate-only (cohort-min=3). | Local neo4j — the client's own graph, full read/write, fast local access, source of truth for their data. |
| Memory | Central atlas-memory — cross-session patterns, cross-tenant signals (privacy-gated). | Local atlas-memory — per-session nodes, private memory. Never leaves client hardware without an explicit sync. |
| Inference | None. Brains is graph + memory only. No model calls run here. | ArcRouter on client hardware — all inference stays local. Brains never sees raw prompts or completions. |
| Sync direction | Receives curated graph writes via Brain API — structured entity nodes and relationships only. | Writes to local graph first. A scheduled sync job pipes entity data up to Brains. Client always has the primary copy. |
| Access model | Tenant-scoped reads only. A client queries only their own slice. No raw cross-tenant access. | Full admin access to their own graph. No write-back path from Brains to client hardware. |
4 — Snapshot Deployment
When a new client is confirmed, the @arcgrid/brain-client SDK provisions their Brain from a snapshot of the main. The client gets a frozen copy of the current entity taxonomy and base knowledge. The main Brain keeps compounding; the client's is their starting baseline.
Export a snapshot from main Brain
BrainClient.snapshot.create() exports the entity graph, schema, and atlas-memory seed data into a versioned bundle. Sensitive data (T3-confidential observations) is stripped before export.Register the client as a tenant in ArcHub
BrainClient.tenants.create({ name, tier, hardware }) provisions a tenant record, issues a sk-... API key, and reserves a database slot on the Brains droplet for their sync endpoint.Deploy the Brain stack on client hardware
The BRAINLIC deployment package installs neo4j (or ArcadeDB), atlas-memory, and the neo4j-mcp server via Docker Compose. The snapshot bundle is imported on first boot. ArcRouter is included for T3 clients; T1/T2 route through a shared endpoint.
Wire the sync job
A scheduled cron runs
brain-sync push --tenant <id> --target brains.arc-grid.com:8050. It diffs the local graph against the last push, exports new entity nodes and relationships, and writes them to the tenant's database on Brains. Raw content never leaves.Ongoing: pull taxonomy updates
The client can pull new entity types and schema migrations from the main Brain via
brain-sync pull --updates-only. This is how we push improvements downstream without overwriting their local knowledge. It is opt-in per release.Key SDK methods
import { BrainClient } from '@arcgrid/brain-client';
const brain = new BrainClient({
apiKey: process.env.BRAIN_API_KEY, // sk-... from ArcHub
baseUrl: process.env.BRAIN_API_URL, // default: http://arcrouter:4000
});
// Create a snapshot for client deployment
const snap = await brain.snapshot.create({
includeSchema: true,
includeBaseEntities: true,
stripConfidential: true,
label: 'star-seeds-v1',
});
// Register a new tenant
const tenant = await brain.tenants.create({
name: 'Star Seeds Co.',
tier: 'T2',
hardware: 'client-managed',
});
// tenant.apiKey → sk-ss-xxxx
// tenant.syncUrl → brains.arc-grid.com:8050/tenants/star-seeds
// Sync local graph entities up to Brains
await brain.sync.push({
tenantId: tenant.id,
since: lastSyncAt,
entityTypes: ['RoadmapTask', 'Client', 'Organization', 'Contact'],
});5 — Client Roster
| Client | Tier | Host | Status | Notes |
|---|---|---|---|---|
| Swift Innovation | T3 | ArcHelm 10.20.10.29 | Partial | neo4j, atlas-memory, ArcRouter, neo4j-mcp all running on ArcHelm. Brains droplet provisioning + sync job still needed. |
| ADage1 | T2 | Client compute | Planned | Full Brain stack (neo4j, ArcRouter, atlas-memory, neo4j-mcp) not yet deployed to client hardware. Brains sync not wired. |
| Funding Forge | T2 | Client compute | Planned | Full Brain stack (neo4j, ArcRouter, atlas-memory, neo4j-mcp) not yet deployed to client hardware. Brains sync not wired. |
| Packly | T2 | Client compute | Planned | Full Brain stack (neo4j, ArcRouter, atlas-memory, neo4j-mcp) not yet deployed to client hardware. Brains sync not wired. |
| Avalon | T3 | Client hardware | Pipeline | Cliff Smith. Hair visualization. vision.hair.transfer. BL-R8 open (biometric compliance gate). |
| Star Seeds (Beyond-Sleep) | TBD | TBD | Pipeline | Star Seeds Co., LTD (Taiwan). Parent of Beyond-Sleep. Current Idea Tavern client. |
| Ag Express | TBD | On-prem | Pipeline | Eric Randolph CEO. Ag-electronics repair platform. 87 features, 4 phases, MVP Q1 2027. |
| Aris | TBD | TBD | Pipeline | Prospect. Pending close. |
| SoulAura | TBD | TBD | Pipeline | Prospect. Pending close. |
6 — Privacy & Data Boundaries
- Inference stays local. ArcRouter runs on client hardware. Brains never sees a raw prompt, completion, or session transcript.
- Entities only, not content. The sync job extracts structured graph nodes and relationships. Raw documents, chat messages, and session logs are never synced up.
- Tenant isolation on Brains. Each client is a separate database. One tenant cannot query another's data — cross-tenant queries are aggregate-only with cohort-min=3.
- No write-back from central. Brains can only receive data from clients. There is no path for the central graph to write to a client's local Brain without a manual pull initiated by the client.
- Confidential data stripped on snapshot. T2/T3-classified observations are excluded from snapshot exports. Clients delivered at sale get only base entity taxonomy and schema.
- Learn-back is opt-in. The main Brain can improve from client graph patterns only if the client's BRAINLIC agreement includes the learn-back clause (BL-R5/R6 — requires counsel review).