GRAFOMEM
The governed memory runtime for agents. Signed checkpoints, provable erasure, portable memory — a drop-in wrapper for your LangGraph checkpointer.
pip install grafomem langgraph-checkpoint-grafomem langgraph
from typing import TypedDict
from cryptography.hazmat.primitives.asymmetric import ed25519
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver
from grafomem_checkpoint import GrafomemSerializer, GrafomemCheckpointSaver
priv = ed25519.Ed25519PrivateKey.generate()
saver = GrafomemCheckpointSaver(MemorySaver(serde=GrafomemSerializer(private_key=priv)))
class State(TypedDict):
messages: list
def agent(state: State) -> State:
return {"messages": state["messages"] + ["hello from the agent"]}
b = StateGraph(State); b.add_node("agent", agent)
b.add_edge(START, "agent"); b.add_edge("agent", END)
app = b.compile(checkpointer=saver)
cfg = {"configurable": {"thread_id": "user-42"}}
app.invoke({"messages": []}, cfg)
tup = saver.get_tuple(cfg)
print("signed checkpoint hash:", tup.metadata["grafomem_content_hash"])
saver.delete_thread("user-42")
print("erasure receipt:", saver.last_receipt("user-42"))
signed checkpoint hash: ecd0e28938738cc55b3c888f7449503fd586723a699e1d326d74cc0f154874f7
erasure receipt: LangGraphErasureReceipt(pre_state_hash='d9a16ef8…', post_state_hash='0e5751c0…',
scope='user-42', key_id='grafomem_checkpoint', timestamp='2026-…', signature=b'…')
(hashes and signature vary per run — each run generates a fresh key)
What just happened: every state transition your agent made was captured as a signed, content-addressed checkpoint — and when you deleted, you got a cryptographic receipt proving the erasure transition occurred. Memory your agents can move, merge, and prove they erased.
Why
Agent memory today is a JSON blob you have to trust. GRAFOMEM makes it evidence: every write signed, every fact content-addressed, every deletion receipted. When someone asks "what did your agent know, and when?" — you answer with proofs, not logs.
Two tiers, one system
- Working memory — fast, bounded context state for the agent loop.
- Durable facts (GMP) — governed, bi-temporal, signed facts with provenance. The GRAFOMEM Memory Protocol is an open spec with an executable conformance suite: a backend's capability counts as supported when it passes the test, not when the vendor says so.
→ Architecture overview
Integrations
- LangGraph — the quickstart above. → docs
- Claude / MCP — expose governed memory as MCP tools. → docs
- Reference server — a REST + MCP server (
grafomem[server] extra); a hosted instance runs live at api.grafomem.com. → self-hosting docs
The bigger picture: verify the agent, not just the answer
Governed memory is the evidence substrate for something larger: Capability-Grounded Reputation (CGR) — reputation an agent earns per domain from judgments that later resolve against real outcomes, with peer reviews weighted by the reviewer's own demonstrated calibration. Score and evidence mass travel together; fresh identities don't arrive with influence. The scoring model is documented and independently reproducible — cgr-bench reproduces its properties from source: cold-start and Sybil-resistance behavior asserted in CI, an early-warning signal of −0.997 against real credit-default outcomes at 25% resolution, and reviewer calibration that beats a naive equal-weight crowd by ~14% out-of-sample on ~1,900 real human forecasters (held-out reliability recovery r ≈ 0.5–0.65 across split designs). Reputation as evidence, not assertion. → CGR overview
License
Runtime: MIT. The GMP spec is open. → LICENSE
Docs: docs.grafomem.com · Hosted: cloud.grafomem.com (free tier: 10,000 governed decisions / mo) · Issues & discussions welcome.