智能体记忆协议

AMP: The Agent Memory Protocol 🧠

The Open Standard for Agentic Memory.

The (Short) Story

I was tired of building AI agents that forgot everything the moment I closed the terminal.

RAG (Retrieval Augmented Generation) is great for documents, but terrible for experience. It chunks text blindly, losing the narrative. When I asked my agents "Why did we decide this yesterday?", they gave me hallucinated nonsense.

So I built AMP. It's not just a database; it's a Hippocampus for your agents. It mimics the human brain's distinction between Working Memory (Short-term context) and episodic Long-Term Memory, giving your agents a continuous, evolving sense of self.

Why developers are switching to AMP?

🌌 Galaxy View (Visualization)

Don't just guess what your agent knows. See it. AMP comes with a local dashboard. Watch memories form constellations in real-time. Nodes cluster by semantic meaning—if two ideas are related, they physically move together.

🕸️ Force Mode (Physics)

Toggle to Force Mode to see the topological connections between your memories. It uses a physics simulation (D3.js) to show you how different memory clusters are "pulled" together by shared context.

🔍 Semantic Query

Stop guessing keywords. Query your agent's memory using natural language. I built a dedicated interface that not only finds relevant memories but shows you the Relevance Score (0-100%) so you know exactly why a memory was retrieved.

🔌 MCP Native (Plug & Play)

Built from day one for the Model Context Protocol.

• Claude Desktop: Add AMP to your config, and Claude remembers you forever.
• Cursor: Give your coding assistant persistent context of your project history.

🧠 The Two-Layer Brain

I don't just dump text into a vector store. I structure it:

1. ⚡ STM (Short Term): High-fidelity buffer. "What are we doing right now?"
2. 📚 LTM (Long Term): Consolidated insights. "What did we learn last week?"

An entity table exists in the schema, but it stores extracted names as nodes only — no edges, no relations — and it is off by default. It is not used by retrieval and contributes nothing to the results below. The relational graph layer is on the Roadmap, not in the box.

LoCoMo Benchmark Results

I evaluated AMP on the LoCoMo multi-session dialogue benchmark: 3 conversations, N = 150 stratified questions per system, seed 42.

The scoring rule, plainly. Answers are graded by two independent local LLM judges (qwen3:8b and deepseek-r1:8b), and accuracy is the mean of the two. Before the judges see anything, any prediction that begins with a refusal phrase ("I don't know," "I cannot determine," …) is scored WRONG — because when shown a refusal against a real ground truth, these judges mark it CORRECT 79–100% of the time. Intervals are 95% bootstrap CIs (10,000 resamples).

How to read this honestly:

• Full Context beats AMP. Paired bootstrap puts AMP-Padded2 10.0 pp behind it, 95% CI (−15.0, −5.0) — a real gap, not noise. If your whole history fits in the context window and you don't mind paying for it, that is the accurate option.
• AMP is statistically tied with naive RAG. +5.7 pp, 95% CI (−0.3, +11.7). The interval crosses zero, so this sample does not separate them.
• What AMP buys you is prompt cost, not a higher score. AMP-Padded2 sends roughly an order of magnitude fewer input tokens per query than Full Context (measured 8.9–16.4× fewer), and it does not need the conversation to fit in the context window at all.
• The neighbor padding is the part that measurably works. w = 2 beats anchor-only retrieval by +11.3 pp, 95% CI (+6.3, +16.7).

Full method, per-category results, latency table and limitations: paper/. Raw per-question verdicts for every system, both judges: benchmarks/results/locomo_local/results.csv. To re-run it yourself: paper/REPRODUCE.md.

Quick Setup

1. Install via uv (Recommended)

# Install the tool
uv tool install amp-memory

# Start the brain
amp serve

2. Or, Install via pip

pip install amp-memory
amp serve

3. Open the Dashboard

Visit http://localhost:8000. The interface is Galaxy Mode by default. Switch to Force Mode to see physics-based connections.

4. Connect to IDEs & Tools

AMP works native with Antigravity, Cursor, VS Code Copilot, and Claude Desktop. Add this to your MCP configuration file (usually mcp_config.json or claude_desktop_config.json):

{
  "mcpServers": {
    "amp-memory": {
      "command": "uv",
      "args": ["tool", "run", "amp-memory", "serve"],
      "env": {
        "PYTHONPATH": "."
      }
    }
  }
}

Now you can say:

"@amp remember that I am refactoring the login controller." "@amp what was the last bug we fixed?"

It knows.

Roadmap 🗺️

• [x] Galaxy View: Visual Semantic Space.
• [x] Graph API: D3.js powered visualization.
• [x] Semantic Search: Vector-based relevance sorting.
• [ ] Entity Graph: Relations between entities — "How is function A related to bug B?" The entity table today stores names as nodes only, with no edges and no co-occurrence weights, and is disabled by default. Edges are future work.
• [ ] Cloud Sync: Sync memories across devices.
• [ ] Multi-Agent Swarm: Shared memory for agent teams.

Star History

Made with ❤️ by Akshay.