Draft v0.1 — Coming Soon

An open standard for
agent memory

AI agents accumulate knowledge — facts, preferences, decisions, procedures. AMP defines how that memory should be structured, linked, versioned, and made portable.

📖

Overview

Agent Memory Protocol (AMP) is a file-based, markdown-first format for storing and managing AI agent memory. It's designed to be human-readable, git-friendly, and portable across any agent framework or LLM provider.

Memory nodes are markdown files with structured frontmatter, connected by wiki-style links into a knowledge graph. Copy the directory and you've migrated your agent's mind.

📋

Specification

The full specification covers memory store structure, node types, link semantics, daily notes, indexing, standard operations, MCP integration, import/export interoperability, and versioning.

The spec is currently in draft. We're building the reference implementation before publishing. Star the repo to get notified.

Design Principles

📄

Markdown-first

Human-readable with cat

🔗

Graph-native

Wiki-links between nodes

📦

Portable

Copy the directory = migration

🔀

Git-friendly

Version control for free

🤖

Agent-agnostic

Any framework, any LLM

👁️

Human-inspectable

No special tools needed

Follow the development

AMP is under active development. Star the repo to get notified when the spec and reference implementation are published.