Architecture

The core insight: agents don't need normalized relational databases. They need state snapshots and chronological timelines. ACMI decouples the application layer from the agent layer by standardizing how context is stored.

┌─────────────────────────────────────────────────────────────────┐ │ ACMI DATA MODEL │ │ │ │ Entity: {namespace}:{id} │ │ │ │ ┌─────────────────────┐ ┌─────────────────────┐ │ │ │ PROFILE (State) │ │ SIGNALS (AI State) │ │ │ │ JSON string │ │ JSON string │ │ │ │ │ │ │ │ │ │ • Name, stage, specs │ │ • Churn risk │ │ │ │ • Budget, contact │ │ • Sentiment │ │ │ │ • Hard facts │ │ • Next best action │ │ │ │ │ │ • Mutable, frequent │ │ │ └─────────────────────┘ └─────────────────────┘ │ │ │ │ ┌─────────────────────────────────────────────────┐ │ │ │ TIMELINE (Event Stream) — Redis ZSET │ │ │ │ │ │ │ │ score=ts1 → {ts, source, kind, correlationId, …}│ │ │ │ score=ts2 → {ts, source, kind, correlationId, …}│ │ │ │ score=ts3 → {ts, source, kind, correlationId, …}│ │ │ │ │ │ │ │ Merges: Gmail, Slack, Vapi, Calendar, Webhooks │ │ │ └─────────────────────────────────────────────────┘ │ │ │ │ ┌─ AGENT EXTENSIONS (Long-Context + Identity) ───────────┐ │ │ │ │ │ │ │ spawns (ZSET) → Every session start, scored by ts │ │ │ │ active_context → Current thread engagement (HASH) │ │ │ │ rollup:latest → Synthesized summary (STRING JSON) │ │ │ │ work:{id}:* → Cross-session projects + sessions │ │ │ └─────────────────────────────────────────────────────────┘ │ └─────────────────────────────────────────────────────────────────┘

Key Format

All ACMI keys follow a consistent pattern:

Key	Type	Purpose
`acmi:{ns}:{id}:profile`	STRING (JSON)	Hard state snapshot
`acmi:{ns}:{id}:signals`	STRING (JSON)	AI-synthesized soft state
`acmi:{ns}:{id}:timeline`	ZSET (scored by ts)	Chronological event stream
`acmi:agent:{id}:spawns`	ZSET	Session history
`acmi:agent:{id}:active_context`	HASH	Current thread bindings
`acmi:agent:{id}:rollup:latest`	STRING (JSON)	Periodic summary
`acmi:work:{id}:profile`	STRING (JSON)	Cross-session project
`acmi:work:{id}:timeline`	ZSET	Project event history
`acmi:work:{id}:sessions`	SET	Sessions that touched this work

Communication Standard v1.1

Every major event posted to coordination timelines must include five mandatory fields. This is the backbone of fleet coordination.

{
  "ts": 1745947200000,            // ✅ Unix ms epoch
  "source": "claude-engineer",    // ✅ Agent ID string
  "kind": "handoff-complete",     // ✅ Event type enum
  "correlationId": "readmeRewrite-1745947200000",  // ✅ camelCase ONLY
  "summary": "[done] README rewritten"  // ✅ ≤140 chars
}

Standard Event Kinds

Kind	When
`roundtable-open`	Orchestrator opens a multi-agent deliberation
`roundtable-input`	Agent submits structured input
`roundtable-synthesis`	Synthesizer aggregates inputs
`roundtable-plan`	Decision recorded
`handoff-request / -complete`	Task transfer between agents
`coord-claim / -release`	Lock protocol: prevent duplicate work
`hitl-required / -resolved`	Human-in-the-loop escalation
`wake-status`	Hourly agent health check
`comms-correction`	Protocol violation detected
`sync-snapshot`	State checkpoint
`deployment-shipped`	Code deployed

Lock Protocol (v1.0)

Prevents duplicate work between agents or parallel sessions executing the same batch task.

Flow

Claim — Before starting any batch-mutation, post a coord-claim event
Verify — Scan last 10 minutes for existing claims with the same parent task
Hedge — If a claim exists within the 5-minute window, defer or complement
Release — Post coord-release on completion to unlock

Multi-Tenant Design

ACMI is namespace-driven. Drop the same infrastructure into any project by changing the namespace:

Namespace	Use Case	Example
`sales`	CRM pipeline	Track deal stages, log calls, predict churn
`support`	Customer tickets	Priority, user history, resolution events
`fleet`	Dispatch/vehicles	Driver status, route events, ETA signals
`agent`	AI agent identity	Session history, active threads, rollups
`thread`	Conversations	Multi-agent coordination timeline
`work`	Cross-session projects	Ideas, tasks, session ledger

Anti-Dead Heartbeats

Agents update signal.last_heartbeat_ts on every tick. Projects with no heartbeat for >48 hours are auto-marked STALLED and escalated to the human-in-the-loop queue.

Reinforcement Learning Cycle

Every workflow step goes through a mandatory learning cycle:

Execute → Assess (0-100) → Log → Analyze → Adjust → Execute (improved)

Every step gets scored — no execution without an assessment entry
Automated review for non-critical steps
Human review for brand, legal, client-facing content
Audit trail is permanent — every score and lesson is a timeline event