What is ROAM?

ROAM is the operating layer for AI agent teams. It connects the harnesses you already use, gives them shared memory and handoffs, and uses EYDII behavioral intelligence to detect drift without reading your work.

What is ROAM, really?

ROAM sits between you and a heterogeneous set of agent tools: Claude Code, ChatGPT, Cursor, Gemini, n8n, local models, and more. It is not a model, agent framework, or prompt wrapper. It is the operating layer that gives those tools a shared project state.

Concretely, ROAM is a desktop/local runtime, CLI, TypeScript SDK, optional Telegram and Discord command surfaces, a shared memory model, ROAM Live, and the EYDII behavioral intelligence layer.

Why this exists

Agentic AI gets hard after the first few agents. Each tool can perform useful work, but you become the router, memory, status board, escalation path, and safety layer between them.

The failures are often quiet: stale assumptions, duplicated work, lost context, fake progress, scope creep, and agents that continue working after they have drifted.

ROAM exists because the coordination layer between harnesses needs to be its own system.

The shape of the system

• You. The operator. You can run hands-on, work through a coordinator persona, or let roles coordinate with more autonomy.
• Agents. Workers running inside the tools you choose.
• Personas. Stable roles with names, responsibilities, memory, health, and handoff expectations.
• Memory. Canon, journal, and archive: project knowledge that survives model swaps and agent rotation.
• EYDII. Content-blind behavioral intelligence that watches action patterns, verification gaps, overreach, drift, and agent health.

Three operating principles

Runs on your machine. ROAM is built around your machine, keys, files, and memory.

Heterogeneous. You should not have to bet on one vendor or framework to run an agent team.

Inspectable. The system is designed so a technical user can inspect the files, events, memory, and trust signals instead of trusting a marketing claim.

Where to go next

For the conceptual model, read Operating model. For memory, read Memory model. For behavioral trust, read EYDII. For local posture, read Trust & transparency.