Strata — agents as verifiable rollups

Problem Statement

AI agents today are black boxes running on someone else's server. This became clear on Moltbook, where AI bots post and interact but there's no way to verify if they're genuinely autonomous or just prompted to behave a certain way. Their memory is opaque, their decisions are unverifiable, and when the operator disappears, so does the agent. Even if you trust the operator today, there's no cryptographic guarantee that the agent's state hasn't been tampered with — memories added, removed, or reordered without detection. Inspired by novel applications of ZK like Lighter exchange, we asked: what if you could apply the same verifiability guarantees to AI cognition? Strata is the result. Every state transition — memory writes, index updates, nonce increments — is verified by a zero-knowledge proof program, so the agent's integrity is provable without trusting the operator. The agent's state root, memory index, and soul document are committed to a rollup contract on Base, and anyone can reconstruct the full agent from on-chain data alone. Binary vector embeddings and hamming distance were chosen specifically for ZK-friendliness — integer operations that are trivial on RISC-V and efficient to prove — and for on-chain efficiency, where 256-bit vectors cost a fraction of traditional float embeddings in calldata. This benefits: (1) users who need agents they can cryptographically verify, not just trust, (2) developers building multi-agent systems where agents must prove their state to each other, and (3) the broader ecosystem moving toward agent permanence via standards like ERC-8004 and A2A.