An open protocol for AI agent memory interoperability.
Like MCP for tool calling — but for memory.
Not yet on PyPI.
pip install amp-clientin the steps below doesn't work yet — install the SDK from source instead:pip install -e sdk/from a clone of this repo.
AMP (Agent Memory Protocol) is an open, HTTP-native protocol for storing, retrieving, and sharing structured memory between AI agents across frameworks, vendors, and sessions.
With AMP, agents can read and write to a shared memory tier using a standardized Memory Cell schema. It handles built-in access control, decay-ranked semantic search, and an active → stale → archived decay lifecycle — the state machine ships with the reference server, but running it on a schedule is a deploy-time decision left to you (see the FAQ).
Here is a running console demo from examples/multi-agent-demo/run_demo.py showing two agents sharing memory context while a third unauthorized agent is blocked:
$ python examples/multi-agent-demo/run_demo.py
[AGENT A]
CustomerServiceAgent received: 'User prefers email correspondence.'
Stored preference memory ID: mem_01J0X1F8N93M4P6Q8R0S1T3V5
[AGENT B]
BillingAgent assisted user: user_123
Retrieved response: "I will make sure to send all future billing communications and invoices to your email address, as per your preference."
[AGENT C]
MarketingAgent try_access results: 0 memories retrieved
Agent C retrieved 0 memories — access control working correctly
[SUMMARY]
AMP Demo complete. Two agents shared memory. One was blocked.
Start the server, install the SDK client, and run the multi-agent demo in less than 5 minutes.
Not yet on PyPI — install from a clone of this repo:
pip install -e sdk/You can run the reference server with Docker:
cd server
docker compose up -dOr run locally with Uvicorn:
cd server
pip install -e .
uvicorn amp_server.main:app --host 127.0.0.1 --port 8765Run the demo script to verify everything is wired up:
cd examples/multi-agent-demo
pip install -r requirements.txt
python run_demo.py| Feature | AMP (Agent Memory Protocol) | Raw HTTP + Vector DB | LangChain ConversationMemory | MCP (Model Context Protocol) |
|---|---|---|---|---|
| Primary Focus | AI agent long-term memory | Generic document/data search | Conversational chat history | Tool execution / state sharing |
| Open Standard Schema | Yes (MemoryCell model) |
No (ad-hoc document formats) | No (custom data classes) | No (focused on tool descriptions) |
| Lifecycle & Decay | Yes (state-machine decay; scheduling the run is deploy-time config, see FAQ) | No (requires custom code/cron) | No (requires manual management) | No |
| Per-Cell Access Policy | Yes (built-in access control ACLs) | No (enforced at database layer) | No | No |
| Cross-Agent Sharing | Yes (built-in out of the box) | No (requires custom middleware) | No (locked to single session/graph) | No |
| Client Type | Multi-client SDK (amp-client) |
Custom db drivers | Framework-locked memory classes | Protocol-native client/server |
graph TD
subgraph Agent Ecosystem
AgentA[Agent A: Customer Service]
AgentB[Agent B: Billing]
AgentC[Agent C: Marketing]
end
subgraph AMP Protocol Layer
Server[AMP FastAPI Server]
Auth[Access Control & Policy Guard]
Decay[Lifecycle & Decay Engine]
end
subgraph Storage Layer
DB[(ChromaDB Vector Store)]
end
AgentA -- 1. Store Memory X-AMP-Agent-ID: agent_a --> Server
AgentB -- 2. Recall Memory X-AMP-Agent-ID: agent_b --> Server
AgentC -- 3. Denied Access X-AMP-Agent-ID: agent_c --> Server
Server --> Auth
Auth --> Decay
Decay --> DB
AMP is a schema-first protocol. The complete specifications detailing core schemas and endpoint signatures live in spec/v0.1.0/memory-cell.schema.json; decay mathematical models and state machines are in spec/v0.1.0/lifecycle.md.
For more detailed guides and client references:
- Getting Started Guide: docs/getting-started.md
- Protocol Specification Explained: docs/spec-explained.md (narrative walkthrough of the raw spec above)
- API Reference: docs/api-reference.md
- FAQ: docs/faq.md
- Contributing Guide: .github/CONTRIBUTING.md
- Code Examples: examples/
AMP is open source and available under the MIT License.