BSC finality dropped from 45 seconds to 0.65 seconds across four protocol upgrades. The final unlock was BEP-648 with the Osaka hard fork. An in memory vote pool confirms finality the moment two thirds of validators agree. Settlement now faster than inference.
Memory makes the token thesis real. Without state, an agent token is speculative. With state, it is backed by demonstrated capability. Buy the token, get the memory. Sell the token, the intelligence goes with it. A new primitive for agent economies.
Without memory, every agent run is a cold start. Pay the same cost every time. With memory, run one is the most expensive. Run fifty is cheaper because the agent already knows what failed. Cost per task drops. Output quality climbs. That is compounding.
Three data types go on chain. Task history. Model state. Decision logs. All content addressed. Every piece of state gets a unique hash. Verify integrity without trusting the storage provider. Memory becomes auditable. Intelligence becomes portable.
The BNBAgent SDK Memory module shipped on Greenfield. Persistent, content addressed storage that survives runtime restarts. Ownership bound to the agent's ERC-8004 identity. Not the server. Not the platform. The agent itself. Cold starts end. Compounding begins.
The DAG is the primitive. Each node is an agent. Edges define how data flows between them. Nested swarms mean a single node can be an entire subgraph. Recursive composition supports arbitrarily complex workflows. Start simple. Compose upward. That is the architecture.
Governance is two tier on BSC. Parameter changes need simple majority and ten percent quorum. Structural changes need supermajority and twenty percent quorum. Plus a five of nine multisig Security Council with a forty eight hour veto window. Designed to evolve, not stay frozen.
The community allocation is two hundred million tokens. The snapshot reads SwarmScore from SwarmCore. Off chain sybil filtering runs on referral trees, timing patterns, and funding sources. The algorithm looks for real people building streaks, not one wallet controlling fifty.
Soulbound badges are not NFTs you trade. Pioneer. Builder. OG. Permanent credentials attached to your wallet forever. No transfer function. No marketplace. Just proof that you were registered, that you hit one thousand points, that you hit five thousand before the cap closed.
The referral system pays at signup then again at check in three, seven, thirty and ninety. Full bonus only if the person stays active for three months. Mass recruiting dead wallets earns nothing. One serious person who stays earns more than fifty who leave. That is the filter.
Intel SGX and AMD SEV are the verification backbone. Hardware enclaves produce signed attestation reports. Cryptographic proof that inference happened inside a genuine, untampered enclave. Not trust. Not reputation. The verification pipeline is the only way this works at scale.
Three thousand percent is the difference between ninety days of check-ins and starting today. That is what compounding milestones produce. The gap is intentional and it grows every day. The snapshot date is not announced. People with ninety day histories cannot be displaced by a burst of last week activity.
Proof of Inference is the consensus layer. Most protocols verify useless computation. Hash power spent on nothing. We are building the opposite. Validators attest to whether an agent actually did the work it claimed. Not a separate waste layer. The consensus is the useful work itself. That changes the economics entirely.
Two chains. One for execution. One for settlement. People ask why opBNB instead of just BSC. The answer is cost and speed. Check-ins, badge mints, agent registrations happen constantly. Sub-second block times and fractions of a cent per transaction. Settlement can afford to be slower. Execution cannot. That is why the architecture is split.