Governed memory, verifiable decisions, and reputation earned from real outcomes.
Your LangGraph agent persists user state. A user invokes their right to be forgotten — can you prove the thread was deleted? Standard checkpointers can't. GRAFOMEM's does:
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) saver.delete_thread("user-123") receipt = saver.last_receipt("user-123") # signed, verifiable proof of the erasure
Every checkpoint signed and tamper-evident; delete_thread yields a signed state-transition receipt — a cryptographic proof the erasure operation occurred. (Proof of the transition — not a claim of media sanitization.) ▶ 60-second demo GIF →
Memory is the implementation; governance is the product. Reputation earned from real outcomes is the third pillar.
A fast, fixed-size in-model state M, read y = Mq. Bounded, portable, mergeable, constant-size. Adaptive state as a movable, forkable, erasable object.
A lossless, queryable, model-agnostic fact store. Auditable, survives model upgrades.
Signed transition receipts on the working tier, cryptographic erasure certificates on the durable tier. Both tiers, one audit surface.
GRAFOMEM ships an executable conformance suite (GMP): "supports capability X" means passes the test for X — two-sided, bootstrap-CI. It's how the durable tier earns trust, and the basis for signed capability reports in GRAFOMEM Cloud.
(This is a capability of the platform — not the pitch. The pitch is the runtime above.)
Conformance proves a memory backend does what it claims. CGR — Capability-Grounded Reputation — proves the same for the agents on top of it: each earns a reputation, per domain, from decisions that later resolve right or wrong. It can't simply be bought or declared — reputation accumulates through resolved outcomes, and peer influence is gated by demonstrated calibration.
A score moves only when a decision is checked against what actually happened — never on assertion, confidence, or volume.
Every score carries its evidence mass, so a lucky first call never reads like an established track record.
Peer review counts only as much as a reviewer's calibration, itself earned on verifiable outcomes; unknown reviewers fail closed, so fresh identities buy nothing.
What is Capability-Grounded Reputation? → · Read the CGR spec →
The runtime: working memory + durable facts + governance primitives. Self-host anywhere, forever free. pip install grafomem · langgraph-checkpoint-grafomem · TS .gfm verifier.
The GRAFOMEM runtime is open source under the MIT license; the GMP spec is open.
Hosted, multi-tenant, governed: decision trail, erasure proofs, governance gateway, regulatory reports, third-party attestation. You point your registry at the cloud — a config change, not a migration.