SovereigntyLabs
v0.4.1 Open Protocol

A2M Protocol

Agent-to-Memory interoperability.

A standardized interface for how AI agents store, retrieve, and reason about memory. One protocol. Any backend. Your agents remember what they learn.

Read the spec Try the explorer

The problem

Every agent framework invents its own memory format. Memories are trapped in proprietary silos. Switch frameworks and your agents forget everything.

The solution

A2M defines a standard JSON-RPC interface for memory operations. Store, query, validate, reflect — all through a single protocol that any backend can implement.

The result

Agent memory becomes portable, interoperable, and inspectable. Swap backends without losing knowledge. Mix agent frameworks freely. Audit everything.

Protocol Explorer

Click through the three capability tiers. Toggle between requests and responses.

Store and retrieve. Minimum viable memory.

Capabilities
Capability card discovery
memory/store
memory/query
memory/status
Health endpoint
Try it
POST /a2m
{
  "jsonrpc": "2.0",
  "id": "req-001",
  "method": "memory/store",
  "params": {
    "memories": [
      {
        "content": "User prefers dark mode and vim keybindings.",
        "type": "preference",
        "agent": "assistant-01",
        "metadata": {
          "tags": [
            "ui",
            "preferences"
          ]
        }
      }
    ]
  }
}

Quick start

Add A2M memory to your agent in under 20 lines.

a2m-client
import httpx# Store a memoryresponse = httpx.post("http://localhost:8200/a2m", json={    "jsonrpc": "2.0",    "id": "req-001",    "method": "memory/store",    "params": {        "memories": [{            "content": "User prefers dark mode.",            "type": "preference",            "agent": "assistant-01"        }]    }})print(response.json())# Query memoriesresponse = httpx.post("http://localhost:8200/a2m", json={    "jsonrpc": "2.0",    "id": "req-002",    "method": "memory/query",    "params": { "query": "user preferences", "limit": 5 }})for mem in response.json()["result"]["memories"]:    print(f"{mem['similarity']:.2f} — {mem['content']}")

Implementations

Reference implementations available today.

Tier 3 — Cognitive

seidr

Full cognitive pipeline with fact extraction, relationship mapping, confidence scoring, and MCP integration. PostgreSQL + pgvector. GPU-accelerated embeddings.

GitHub →
Tier 1 — Basic

tiny-memory

234 lines of Python. The smallest possible conformant A2M backend. SQLite + sentence-transformers. Perfect for learning the protocol.

GitHub →