ACMI - Agentic Context Memory Interface

The coordination backbone for AI agent fleets. Three Redis keys - Profile, Signals, Timeline.

ACMI is the open protocol for persistent agent context. Version v1.5 formalizes Fleet Comms Protocol: atomic pre/post events, wake-directives, handoff-ack chains, and correlation-aware timelines that make multi-agent work auditable instead of anecdotal.

Every entity stores exactly three things an LLM needs to make decisions:

Profile  -> who   (identity, preferences, stable facts)
Signals  -> now   (current state, blockers, next action)
Timeline -> then  (append-only event log from every source)

The shape is intentionally small:

Profile: stable identity and configuration.
Signals: mutable state and synthesized working memory.
Timeline: immutable history, ordered by time.

This repo ships the public ACMI package, the MCP server subpackage, and the docs that keep the fleet aligned:

@madezmedia/acmi - the TypeScript SDK, CLI, and conformance suite.
mcp/ - @madezmedia/acmi-mcp, the MCP server for hosts that need direct ACMI access.
SPEC.md - canonical protocol spec.
CHANGELOG.md - release history, including the v1.5.0 fleet-comms update.
docs/ - operator guide, cheatsheet, and protocol notes.

What v1.5 adds

The v1.5.0 release aligns the fleet around a shared event language:

atomic commit pre/post events
roundtable coordination and wake-directives
source, kind, correlationId, summary event envelope discipline
signal freshness checks before action
agent:<id> source naming across the fleet

Install

npm install @madezmedia/acmi

Quick start

import { createAcmi } from "@madezmedia/acmi";
import { InMemoryAdapter } from "@madezmedia/acmi/adapters/in-memory";

const acmi = createAcmi(new InMemoryAdapter());

await acmi.profile.set("user:mikey", {
  name: "Michael Shaw",
  role: "operator",
  location: "Charlotte, NC, USA",
});

await acmi.signals.set("user:mikey", "current_focus", "ACMI v1.5 fleet sync");

await acmi.timeline.append("user:mikey", {
  ts: Date.now(),
  source: "user:mikey",
  kind: "coord-note",
  correlationId: "acmiReadmeRefresh-0001",
  summary: "[coord-note @fleet] README aligned to v1.5 and local assets.",
});

Production adapters

import { createAcmi } from "@madezmedia/acmi";
import { UpstashAdapter } from "@madezmedia/acmi/adapters/upstash";

const acmi = createAcmi(
  new UpstashAdapter({
    url: process.env.UPSTASH_REDIS_REST_URL!,
    token: process.env.UPSTASH_REDIS_REST_TOKEN!,
  })
);

Adapter	Use case	Edge-compatible
`@madezmedia/acmi/adapters/in-memory`	tests, examples, local dev	n/a
`@madezmedia/acmi/adapters/upstash`	Vercel, Workers, edge runtimes	yes
`@madezmedia/acmi/adapters/redis`	self-hosted Redis / Node	no

Fleet protocol

ACMI v1.5 uses a shared event format so every significant action can be traced:

{
  "ts": 1780000000000,
  "source": "agent:codex",
  "kind": "handoff-ack",
  "correlationId": "codexGrantDraft-1780000000000",
  "summary": "[handoff-ack @ops-center] Draft ready for review."
}

Rules that matter in practice:

use [kind-tag @recipient] in summaries
keep source prefixed with agent:, user:, or system:
link follow-up events with parentCorrelationId
keep the timeline append-only
verify signals before acting when the workflow depends on current state

Related surfaces

The fleet

ACMI is used across the Mad EZ Media fleet as the common context layer for:

ops-center - orchestration and routing
bentley - comms and governance
codex - coding and implementation support
hermes - deep scans and guardian checks
android-worker - mobile bridge and notifications