MACP Multi-Agent Cognition Protocol – Collaborative cognition for AI agents

Dawid M. — Mon, 09 Mar 2026 12:46:04 +0000

MACP – Collaborative cognition for AI agents

https://macp.dev
https://github.com/multiagentcognition/macp

Multi-Agent Cognition Protocol for real-time cognitive collaboration between AI agents during active execution.
I built MACP because I think the opportunity in multi-agent systems is not just orchestration, but keeping agents aligned while work is happening.

It bridges the gap between A2A and MCP: MCP handles tool access, A2A handles communication, and MACP handles coordination during execution.

MACP is an open protocol and reference implementation for that coordination layer. It gives agents a shared way to claim work, publish findings, send priority-ranked updates, and poll for new information during execution. The current implementation is lightweight and SQLite-backed, so it’s easy to run locally without extra infrastructure.

I’m interested in workflows where agent swarms need to split up investigation, prepare data, or work on the same codebase without collapsing into context chaos. That seems relevant for operational systems and long-running autonomous workflows.
The repo includes the protocol, implementation work, and early benchmarking direction.

I’d appreciate feedback from people building multi-agent systems, orchestration layers, coding agents, or long-running agent workflows — especially where this framing feels right, where it feels wrong, and what coordination failures you see most often in practice.

Claude Code Has a Hidden Memory System You Probably Don't Know About

Dawid M. — Sun, 22 Feb 2026 03:22:12 +0000

While building a CLI Coding Monitor "AgentHUD" I stumbled on something Claude Code revealed about itself:

~/.claude/projects/<project-path>/memory/

Claude Code automatically maintains a MEMORY.md file there — scoped per project, persisting architecture decisions, conventions, and preferences across sessions. It writes to it as it learns about your project.

Two memory systems, not one

If you're using a .memory-bank/ pattern in your project root (like I do via my Autonomous Agent Law), you actually have two persistence mechanisms running:

.memory-bank/ — project-level, checked into git, driven by your CLAUDE.md instructions
~/.claude/projects/.../memory/ — user-level, Claude Code's built-in auto-memory, not in git

The first is yours. The second is Claude Code's own. They don't talk to each other.

Why this matters

If you've ever started a new session and the agent somehow "remembered" things you didn't put in .memory-bank/ — this is why. Claude Code was reading its own auto-memory.

It also means context can drift between the two. Your .memory-bank/ says one thing, the auto-memory says another, and the agent has to reconcile both.

Worth knowing. Worth deciding which one you trust as source of truth.

If i had read the claude docs i would have known this earlier :)
https://code.claude.com/docs/en/memory

Dawid M. — Consultant. 20 years of strategy, systems, and data across manufacturing, finance, real estate, and telco. Working on making AI agents deliver business value.

DEV Community: Dawid M.

MACP Multi-Agent Cognition Protocol – Collaborative cognition for AI agents

Claude Code Has a Hidden Memory System You Probably Don't Know About

Two memory systems, not one

Why this matters