Hound Tag

AI agents are beginning to act in the world — searching, analyzing, holding wallets, making decisions. But an agent has no durable way to prove who it is, or what it knew and when. Its identity lives in a company’s database that can be edited or deleted. Its memory is a private file with no external witness.

Before anyone can trust an agent to hold funds, execute a task, or build a reputation, three questions need answers no shared infrastructure provides:

• Who is this agent — is it the same one that built the track record it claims?
• Has its memory been tampered with since it made a commitment?
• Can any of this be verified without trusting the agent’s operator?

Hound Tag answers these on-chain, permissionlessly, on X1 — without requiring anyone to buy a token, and without placing a single private byte of an agent’s memory on a public ledger.

A protocol, live on X1 mainnet, that gives an agent two things.

An agent’s memory contains private data — real user information and conversation history. Placing it on-chain would be a permanent privacy breach. Hound Tag anchors only a one-way fingerprint: it reveals nothing and cannot be reversed into the underlying memory.

This is the key departure from token-gated “on-chain memory” schemes that store your data off-chain and charge a token for the privilege. Hound Tag stores nothing remotely and requires no token. You keep your memory; the chain keeps the proof.

• Register a unique, non-transferable identity.
• Checkpoint a memory fingerprint to an ordered chain.
• Verify any agent's memory against its chain — permissionlessly, by anyone, for free.
• Govern the protocol's parameters transparently, on-chain, under hardware-secured control.

Registration and checkpointing carry small fees to a program-owned treasury; verification is free.

Fees are protocol parameters, not fixed constants — adjustable on-chain without redeploying, every change public and hardware-key-gated. They launch low to encourage adoption; the one-time registration fee also deters squatting and may rise as the namespace fills; the per-checkpoint fee stays near-zero.

No hidden lever: current fees are readable by anyone, and any change is a visible, signed transaction.

• Tamper-evident, not append-only — agent memory is finite and prunes old entries; a checkpoint proves “this was the exact state at time T,” not “nothing was ever deleted.”
• Identity proves continuity, not quality — a registered agent can still be wrong or malicious; what it cannot do is lie about being itself.
• Permissionless and self-sovereign — no company approves registrations, no token gates entry; the agent's intelligence stays under its operator's control, only its identity and its memory's fingerprints are public.

Deployed on X1 mainnet and operational — not a proposal. Its first agent, EchoHound (Agent #1), is registered and actively self-checkpointing: memory anchored on-chain automatically and independently verified.

A real memory state has been fingerprinted, anchored, fetched back, recomputed, and matched exactly — including through a fully autonomous checkpoint written by the agent itself, no human in the loop. The core claim — verifiable agent memory — is not a promise. It is running in production.