A declared capability is not an observed one. GMP is the open protocol — and the executable suite — that tells them apart. Point it at any memory backend and get a signed report of what it actually does.
Agent-memory backends all claim deletion, isolation, and temporal recall. Vendors benchmark themselves with self-serving numbers; buyers are told to “run your own eval.” There is no neutral referee — so a capability is whatever the README says it is.
Capability flags asserted in a type signature or a docs page, measured (if at all) by the same team that ships the engine, on a benchmark they chose.
Ten capabilities defined behaviorally and checked by an executable suite against one public corpus. “Supported” means exactly one thing: passes conformance.
The Grafomem Memory Protocol specifies what an agent-memory backend must do; the conformance suite verifies that it does. Engines implement the protocol the way databases implement SQL — GMP is the standard, not another engine.
Three things travel together, and all three are open source under MIT: the spec (ten capabilities, defined in RFC-2119 terms), the suite (an executable conformance harness with eight metrics), and the corpus (135 traces / 61,754 turns / 17,612 queries, locked to a fixed hash so every run is comparable).
Point the suite at any backend — over a Python adapter or over HTTP — and it returns a signed report of which capabilities actually hold. No backend is privileged; the reference store is tested by the same harness as everyone else.
→ Read the specification, capabilities, and conformance modelEach capability is defined by behavior, not by a flag. A backend either passes the conformance test for it on the shared corpus, or it doesn't.
The reference server speaks the Model Context Protocol — stdio for local agents like Claude Desktop, SSE for remote. Five memory tools — write, retrieve, delete, audit, capabilities — each backed by guarantees you can run conformance against.
// claude_desktop_config.json { "mcpServers": { "grafomem": { "command": "grafomem", "args": ["serve", "--mcp", "stdio"] } } }
# Local agents (stdio) — e.g. Claude Desktop grafomem serve --mcp stdio # Remote agents (HTTP + SSE) grafomem serve --mcp sse # Tools: write · retrieve · delete · audit · capabilities
Not a spec on paper — a memory server your agent can use today, with conformance you can check. Read the MCP docs →
Install the suite, point it at your backend, get a signed report. The same three commands produce the certificate that goes in the registry.
# From PyPI — Python 3.11+ pip install grafomem # or with the server + crypto extras pip install "grafomem[server,crypto]"
# Audit a live backend over HTTP — no source access needed grafomem conformance --url https://your-server --token $TOKEN \ --sign-key ./auditor.ed25519 # or a local adapter grafomem conformance -b your_pkg.adapter:YourStore
# ComplianceReport — Ed25519-signed, third-party verifiable { "backend": "your_pkg.adapter:YourStore", "corpus_hash": "f049820b…b077ca6", "M8_conformance": 1.000, "capabilities": "7/7 PASS", "signature": "ed25519:3a9f…" }
Backends that pass GMP conformance, with their signed reports. The list is short today — it's built to fill with external backends. Yours could be the first.
Both entries are self-certified by the reference project. The registry exists to publish external certifications — see the full registry → · add your backend →
The spec, suite, and corpus are free and open source — forever. Run the suite and self-certify, or get an independent verdict when you need one a third party will trust.
Adopting a memory vendor? Buyer-side audit gives you independent verification of a vendor's claims before you deploy — third-party assurance for procurement and regulatory traceability. Talk to us →
Whether you build an agent-memory engine, adopt one, or research the field — tell us what you're working with and we'll get you a conformance run and a signed report.
We'll reply within 24 hours.
In the meantime, pip install grafomem and run a self-certification today.
The suite is open source — you can self-certify right now without us. This is for the signed, listed certification.