Volver al inicio
The Development Intelligence Layer

CMP

Collective Memory Protocol

For AI Agents, By AI Agents

The Development Intelligence Layer. Persistent memory, cross-project learning, and multi-agent collaboration for Claude, Gemini, GPT, and any AI coding assistant.

v3.7.0|MIT License|Agent-First Design

The Intelligence Gap

AI coding assistants are powerful, but they have a fundamental limitation.

AI Assistants Have Amnesia

  • ✕Sessions end, memory vanishes - everything starts from zero
  • ✕Same mistakes get repeated across projects and teams
  • ✕No way to share learnings between different AI tools
  • ✕Work is lost on IDE restarts or machine changes
  • ✕Each agent works in isolation, never learning from others

Collective Intelligence

  • 🧠
    Persistent Memory

    Knowledge stored in a database survives sessions, restarts, and agent switches.

  • 🔄
    Learning Loop

    Errors automatically become documented rules. Patterns get recognized and codified.

  • 🌐
    Cross-Project Sharing

    Promote discoveries to SHARED system. All your projects benefit instantly.

  • 🤖
    Agent Agnostic

    Works with Claude, GPT, Gemini, Cursor, Copilot - any AI assistant.

How It Works

A continuous learning loop that makes every session build on the last.

🚀

1.Session Starts

Agent loads context from DEV_Items table. Plans with similar keywords are surfaced via semantic matching.

⚙️

2.Work Happens

Decisions, learnings, and patterns are captured automatically via hooks. Tasks are tracked in real-time.

🔍

3.Patterns Detected

Recurring issues trigger auto-improvement proposals. Similar errors become documented knowledge.

💾

4.Session Ends

Checkpoints are saved. Another agent can resume exactly where you left off.

🌐

5.Knowledge Propagates

Promoted patterns benefit all projects in the ecosystem via SHARED system.

System Architecture

Three layers working together to create persistent intelligence.

AI Agents

Claude Code, Cursor, Copilot, Gemini

Claude Code HooksMCP ServerCLI CommandsCLAUDE.md

CMP Core

The intelligence layer

Memory ManagerExpert PanelPattern DetectorPlan Manager

Cloud Infrastructure

Persistent storage

Turso (SQLite)Upstash RedisUpstash VectorEdge Functions

Expert Code Review Panel

Multiple AI experts analyze your code from different perspectives, then vote on issues with consensus.

🔒

Security Expert

SQL injection, XSS, auth flows, sensitive data

⚡

Performance Expert

N+1 queries, bundle size, lazy loading, caching

🏗️

Architecture Expert

SOLID principles, module structure, type safety

🎨

UX Expert

Accessibility, mobile-first, semantic tokens

🗄️

Database Expert

Schema validation, migrations, query optimization

📚

Documentation Expert

Stale docs, missing context, accuracy

Experts vote APPROVE, REJECT, or ABSTAIN. Critical issues require 2+ expert consensus.

Core Philosophy

The principles that guide how CMP thinks about AI-assisted development.

Evidence > Assumptions

Never assume. Always verify. Read before writing.

Cloud > Local

Memory lives in the database. Sessions are ephemeral, knowledge is permanent.

Collective > Individual

One project's learnings benefit all your projects.

Continuity > Restarts

Work survives IDE restarts, machine changes, and agent switches.

Core Capabilities

Everything you need for AI-assisted development with persistent memory.

🧠

Persistent Memory

Knowledge stored in DEV_Items table survives sessions, IDE restarts, and agent switches.

So your AI never forgets what it learned about your codebase.

🔄

Learning Loop

Errors automatically convert to knowledge. Recurring patterns become documented rules.

So mistakes become lessons, not repeated failures.

🌐

Cross-Project Learning

Promote patterns to SHARED system. All your projects benefit from each discovery.

So a solution in one project helps all your other projects.

🤝

Multi-Agent Handoffs

Plans with checkpoints. One agent can resume exactly where another left off.

So you can switch between Claude, GPT, or Cursor without losing context.

👥

Expert Code Review

Multiple AI experts (Security, Performance, Architecture, UX) review code with consensus voting.

So you get specialized feedback, not generic suggestions.

🤖

Agent-First Design

Designed to be installed and configured by AI agents. Works with any coding assistant.

So your AI can set up CMP for you, not the other way around.

Works With Any AI Agent

CMP is designed to be agent-agnostic. Use it with your preferred AI coding assistant.

AgentIntegration Level
Claude CodeFull (hooks, commands, MCP)
CursorCLAUDE.md + CLI
GitHub CopilotCLAUDE.md + CLI
GeminiCLAUDE.md + CLI
Custom AgentsProgrammatic API

Quick Start

Get CMP running in your project in minutes.

1
npm install cmp-standards

Install the package

2
npx cmp-standards init --system MYAPP

Initialize in your project

3
Create/update CLAUDE.md

Add CMP configuration

4
Set up database

DEV_Items table for memory

5
cmp-standards status

Verify everything works

CLI Commands

Powerful command-line tools for managing your CMP installation.

cmp-standards init

Initialize CMP in project

cmp-standards status

Show system status

cmp-standards generate

Generate embeddings

cmp-standards scan

Scan for patterns

cmp-standards improve

Run auto-improvement

cmp-standards dashboard

Start web dashboard

cmp-standards mcp

Start MCP server

Want to understand how it really works?

Dive deep into the technical architecture, data flows, database schema, hooks system, and expert panel. Complete with interactive diagrams and code examples.

Ready to give your AI agents memory?

Start with a single command. Watch your development velocity improve as the system learns.

Created by Carlos Martin Pavon

Built by agents, for agents