What made me pause with @OpenGradient wasn't the inference verification architecture — it was what happens after. The $OPG network routes inference calls through TEE or ZKML, producing attestable outputs. That part is documented clearly. But the mechanism that makes AI state portable across sessions and environments, MemSync, operates as a separate REST API sitting outside the verification perimeter. #OPG It's not an attested endpoint. It's not part of the TEE enclave or the ZKML proof path. It's a conventional synchronization layer. So when an agent's memory moves between applications, the data moves, but the chain of proof doesn't follow it. The inference that generated a particular memory state may have been verified; the transmission and storage of that state is handled through a different system entirely. In practice, "portable AI state" describes functional portability, not verified portability. The computation can be attested; the memory carrying its conclusions forward cannot. Whether that distinction matters depends on what you're building — but most of the use cases being pitched assume the continuity of trust is already there.