AMP – Agent Memory Protocol for AI Agents

TL;DR

TL;DR: AMP is an open-source memory server for AI agents that gives LLMs a persistent, structured hippocampus — with MCP support, a Galaxy View visualization, and a 3-layer memory architecture (STM, LTM, Graph).

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: ampmemory.io — verified via GitHub README
• Source repository: github.com/akshayaggarwal99/amp — README read end-to-end
• License: MIT (verified via LICENSE badge in README)
• HN launch thread: news.ycombinator.com/item?id=46259498
• Source last checked: 2026-07-09 (commit main, pushed 2026-05-13)

What Is AMP?

AMP stands for Agent Memory Protocol — an open-source memory server that gives AI agents a persistent, structured sense of self. Built by akshayaggarwal99, it is not a vector database with RAG dumped on top; it mimics the human brain’s distinction between working memory and long-term episodic memory.

The README describes it as giving LLMs a hippocampus — a dedicated organ for experience, not just document retrieval.

Core Architecture: The 3-Layer Brain

AMP organizes memory across three layers:

1. STM (Short-Term Memory) — High-fidelity buffer for immediate context. The equivalent of “what are we doing right now?”
2. LTM (Long-Term Memory) — Consolidated insights and episodic summaries. The equivalent of “what did we learn last week?”
3. Graph — Entity connections between memories. Shows how function A relates to bug B, modeled with D3.js force-directed graphs.

Galaxy View Visualization

One of AMP’s most distinctive features is Galaxy View — a 60fps local dashboard where memories form constellations in real-time. Related nodes physically cluster together by semantic meaning. You can also toggle Force Mode to run a physics simulation that shows topological connections between memory clusters.

Benchmarks Against Mem0

The author benchmarked AMP against Mem0 (the leading competitor) on the complex LoCoMo dataset:

| System | LLM Recall Accuracy | Why? | |---|---|---| | AMP | 81.6% | Context-First — preserves the narrative | | Mem0 | 21.7% | Extraction-First — aggressive summarization loses detail |

The gap is large because AMP is context-first, not extraction-first.

Setup Workflow

Prerequisites

• Python 3.10 or higher
• uv package manager (recommended) or pip

Step 1: Install AMP

uv tool install amp-memory

Or via pip:

pip install amp-memory

Step 2: Start the Memory Server

amp serve

This starts the local server with Galaxy View accessible in your browser.

Step 3: Connect to an MCP Client

AMP is MCP-native (Model Context Protocol). Add it to your MCP configuration:

{
  "mcpServers": {
    "amp": {
      "command": "amp",
      "args": ["serve"]
    }
  }
}

Works with Claude Desktop, Cursor, and any other MCP-compatible client.

Deeper Analysis

Why Memory Matters for AI Agents

Standard RAG (Retrieval Augmented Generation) chunks text blindly and loses narrative context. When you ask an agent “Why did we decide this yesterday?”, a RAG pipeline retrieves semi-relevant documents but cannot reconstruct the experience of the decision. AMP solves this by preserving episodic structure — the story of what happened, not just the facts.

MCP-Native Design

Building on the Model Context Protocol means AMP works out-of-the-box with Claude Desktop, Cursor, and a growing ecosystem of MCP-compatible tools. There is no custom integration layer required — if your tool speaks MCP, AMP is a drop-in memory backend.

Visualization as a Debugging Tool

Galaxy View is not cosmetic — it is a real-time window into what your agent knows and how that knowledge is structured. Watching memory nodes cluster by semantic meaning lets you catch gaps or false connections early.

Practical Evaluation Checklist

• [ ] Installs via uv tool install amp-memory without errors
• [ ] amp serve starts the local server
• [ ] Galaxy View loads in browser at the reported port
• [ ] MCP client connects and persists context across sessions
• [ ] Semantic query returns relevant memories with relevance scores
• [ ] Force Mode physics simulation renders correctly

Security Notes

• All memory is stored locally — no cloud dependency for the core server
• The MCP connection is local by default; secure it behind your own network boundary for production use
• No authentication layer is documented in the README — treat it as an internal tool

FAQ

Q: Does AMP require an external database? A: No. AMP uses SQLite for persistence and D3.js for visualization — both bundled. No external services needed.

Q: How is this different from a vector database with RAG? A: RAG is extraction-first — it chunks text and loses narrative. AMP is context-first and preserves episodic structure, achieving 81.6% recall on LoCoMo vs Mem0’s 21.7%. The 3-layer brain model (STM, LTM, Graph) is specifically designed for agentic experience, not document retrieval.

Q: Which MCP clients are supported? A: Claude Desktop and Cursor are explicitly mentioned. Any MCP-compatible client should work since AMP implements the standard Model Context Protocol.

Q: Can I self-host this? A: Yes. The server runs locally via amp serve with no cloud dependency. Everything stays on your machine.

Conclusion

AMP fills a real gap in the AI agent stack: agents that remember. Its 3-layer memory architecture, Galaxy View visualization, and MCP-native design make it a compelling open-source alternative to Mem0 for developers who want persistent, structured agent memory without a cloud dependency. If you are building agents that need continuity of experience across sessions, AMP is worth a look.

Install it in 30 seconds with uv tool install amp-memory and run amp serve to start.