Overview
The neo4j-memory MCP server is the programmatic interface to the Grid's knowledge graph. All project facts, roadmap tasks, client records, people, decisions, and platform entity relations live in neo4j. Agents and Claude Code sessions interact with the graph exclusively through this MCP — never via direct Cypher connections.
This is distinct from atlas-memory MCP (port 8030), which handles per-session and per-user memory with ONNX embeddings. neo4j-memory is the canonical long-term knowledge graph.
Connection
# Claude Code MCP config (~/.claude/settings.json or project .mcp.json)
{
"mcpServers": {
"neo4j-memory": {
"url": "http://localhost:8050/mcp/"
}
}
}The server runs at http://localhost:8050/mcp/ on ArcHelm (Brain node). It is always running as part of the Brain cognitive stack in the agent-space-co/brain repo.
Tools (9)
Graph schema
Key entity types in the Grid graph:
RoadmapTask- linked to"Cost Optimization Roadmap"viaPART_OF; sequential tasks linked viaBLOCKSPerson- team members, clients, partnersProduct- ArcGrid products and modulesProject- active client or internal projectsDecision- ratified architectural or product decisionsIntegration- external services and how they connect
Naming conventions
All new facts — board names, group names, people, templates, URLs, IDs, project context — belong in neo4j, not in CLAUDE.md or system prompts. System prompts hold behavioral rules; the graph holds knowledge. When onboarding a new integration or agent, ask: "Is this a rule or a fact?" Rules go in prompts. Facts go in the graph.