The part that stayed with me about MemSync wasn't the feature — it was where it lives in the stack. @OpenGradient builds $OPG around verifiable inference: TEE enclaves, ZKML proofs, settlement modes that let you choose how much trust you need before a result is accepted. #OPG 's pitch is auditable compute. But MemSync — the layer responsible for giving AI agents continuity across sessions, for making a model behave less like a stateless function and more like something that remembers — sits outside that verification perimeter. The inference gets attested. The memory feeding it doesn't. So you can verify what the agent concluded, but not the accumulated context it was reasoning from when it got there. That's not a bug exactly — persisting memory with the same attestation overhead as inference would be architecturally brutal, and maybe the gap closes later. But it means the trust boundary in this system stops at the compute layer, not at the full decision process. I keep thinking about what "verifiable AI" means when the thing being verified is only the last step.