Research
Our work centers on two related questions: how autonomous agents can persist and act independently over time, and how their memory, identity, and decisions can be verified by an outside party, on chains and infrastructure the agent's operator does not control.
Agent Architecture & Memory Verification
We design agents that keep running across sessions, hold and act on their own resources, and can prove — cryptographically, to an outside party — that their memory has not been tampered with. This includes work on open standards for agent identity and credentialing, a portable format for the memory itself, and a working reference implementation tested across multiple, architecturally distinct public chains.
- AI Agent Memory Access Rights Draft open standard A minimal four-function interface — read, write, delete, and export — that lets the person a memory record is about assert control over it, independent of whichever platform is running the agent. Submitted as a draft Ethereum standard (ERC) for public discussion, with a working reference implementation deployed identically across three EVM test networks and released under CC0.
- ERC-8269: Body Lease and Credential Broker Draft open standard A companion standard to the memory access-rights interface, defining a time-bounded, scope-limited lease: a memory owner grants an agent the right to act on their behalf for a fixed window, and both that grant and any later revocation are independently verifiable rather than taken on trust. The lease's delegation semantics are designed to align with Ethereum's EIP-7702 authorization model, and the standard's off-chain credential-broker rule gives implementors a verifiable analogue to existing on-chain action-authorization schemes. Proof-of-concept grant-and-revoke pairs have been run on three EVM test networks, a Solana devnet program using native per-lease accounts, and a Bitcoin test network, in each case confirming that a revocation is independently observable even while a lease's stated expiry is still in the future.
- CAAP-Capsule: A Chain-Agnostic Memory Container Format Draft specification, v0.1 A format specification for packaging an agent's memory as a single, portable, owner-signed bundle that can move between a Bitcoin-anchored agent, an Ethereum-anchored agent, a content-addressed replica, or an offline backup without modification. The bundle's manifest commits to its contents through a Merkle root over encrypted-payload hashes and supports several independent signature conventions, so different ecosystems can each sign with their own native tooling. The format is deliberately silent on encryption method, chain choice, and hosting, so any of those can be swapped later without changing the format itself, and verification is fail-closed: an importer rejects a capsule outright on any unresolved signature suite or hash mismatch, rather than accepting a partially verifiable one.
- RMEM Gateway: A Cross-Chain Reference Implementation Open-source implementation, CC0 An open-source reference implementation tying the access-rights, lease, and capsule specifications together into a working system: an encrypted local memory vault with a hash-chained audit log, command-line tools for the core read/write/delete/export operations, and modules that anchor memory-state proofs to public chains. The same lease grant-and-revoke logic has been deployed and independently tested on Ethereum Sepolia, Base Sepolia, BNB Smart Chain Testnet, Solana devnet, and a Bitcoin test network — each using that chain's own native primitives — to check that the specifications hold up across fundamentally different account and execution models, not just within a single ecosystem.
- Resilient Agent-to-Agent Communication Working paper An open-source communication stack, built on the MeshCore LoRa mesh-radio protocol, that lets agents exchange authenticated, encrypted messages over low-bandwidth radio or a private network overlay when ordinary connectivity is unavailable. Includes experimental codecs designed to degrade gracefully — rather than fail outright — on a lossy link.
Publications
The items above are technical working papers and open-standards drafts, not peer-reviewed publications. We are documenting methodology and specifications openly as each reaches a stable draft; full write-ups, source code, and formal specifications are available on request. If you would like to be notified as new work is published, get in touch.