grafomem 0.4.0

GRAFOMEM — agent-memory conformance benchmark and compliance toolkit

GRAFOMEM

The governed memory runtime for agents.

Give your AI agents memory they can move, merge, and prove they erased — a fast in-model working state plus a durable fact store, signed and auditable at the substrate.

Fair Source — source-available. The GRAFOMEM runtime core is licensed FSL-1.1-Apache-2.0 (self-host, use, and modify freely; only a competing managed service is restricted; converts to Apache-2.0 after 2 years). The spec, .gfm format, conformance suite, and adapters are Apache-2.0. Please do not describe the runtime core as "open source" — it is Fair Source. (See LICENSE files.)

What it is

A two-tier memory system with governance built in:

• Working memory (CSO) — a fast, fixed-size in-model state M, read y = Mq: bounded, portable, mergeable, constant-size. Adaptive state as a movable, forkable, erasable object.
• Durable facts (GMP) — a lossless, queryable, model-agnostic fact store.
• Governance primitives — signed transition receipts on the working tier, cryptographic erasure certificates on the durable tier.

Quickstart (5 minutes) — governed checkpoints for LangGraph

pip install grafomem langgraph-checkpoint-grafomem

from langgraph.checkpoint.memory import MemorySaver
from grafomem_checkpoint import GrafomemSerializer, GrafomemCheckpointSaver

saver = GrafomemCheckpointSaver(MemorySaver(serde=GrafomemSerializer(private_key, key_id, trust, consent)))
graph = builder.compile(checkpointer=saver)

# ... run your agent ...
saver.delete_thread("thread-123")          # right-to-erasure
receipt = saver.last_receipt("thread-123") # signed, verifiable proof the erasure transition occurred

Every checkpoint is signed and tamper-evident; delete_thread yields a signed state-transition receipt — a cryptographic, verifiable proof that the erasure operation occurred (not a claim of media sanitization).

---

The Protocol (GMP)

GRAFOMEM began as a benchmark for one question — what should a standard for agent memory actually specify? — and turned into the answer: a benchmark, a protocol (GMP), an executable conformance suite, and a certified, persistent reference implementation.

Memory capabilities are orthogonal, a declared capability is not the same as observed behavior, and the only way to tell them apart is to test — so agent memory should be specified and conformance-checked like any other protocol.

The benchmark demonstrates three key results:

1. Four orthogonal axes. Representational capability (versioning / supersession), embedding quality, retention policy, and a two-sided privacy primitive (deletion and tenant isolation) must be specified and verifiable separately.
2. Claims ≠ behavior. A backend can declare HARD_DELETE or MULTI_TENANT and still leak forbidden data.
3. Protocol + conformance. "Supports capability X" is defined operationally: passes the conformance suite for X.

Repository layout

src/grafomem/   working memory runtime (CSO · sdk · mcp · server)   [FSL-1.1-Apache-2.0]
src/aml/        durable facts runtime (GMP · backends)               [MIT]
spec/           .gfm format + SPEC                                   [Apache-2.0]
tests/runtime/  test suite for the working memory tier               [Apache-2.0]
tests/          test suite for the durable tier & GMP benchmark      [Apache-2.0]
adapters/       framework adapters (langgraph-checkpoint-grafomem)   [Apache-2.0]
verifier-ts/    zero-dependency TypeScript .gfm verifier             [Apache-2.0]
examples/       runnable demos                                       [Apache-2.0]

Develop

# Editable install for the entire unified runtime
pip install -e ".[runtime]"
bash ci.sh          # pytest · conformance · TS verifier · Fair-Source lint

Cloud

Self-host everything above for free, forever. GRAFOMEM Cloud adds the hosted, multi-tenant, governed, attested platform — talk to us. You point your registry at the cloud; it's a config change, not a migration.

---

GRAFOMEM is a project of Ulissy s.r.l.