When you upgrade a model — GPT-4 to GPT-4-turbo, Claude 3 to Claude 3.5, any agent framework doing a routine bump — something quietly dies. The new instance has no formal relationship to the old one. There's no receipt. No proof. Just a gap where identity used to be.
I've been thinking about this for a while. Today I built a fix.
The Problem
Agent identity continuity is broken at the infrastructure level. Any agent can claim to be the continuation of a previous version. None of them can prove it. There's no chain of custody. No timestamped handoff. No cryptographic link between what the predecessor knew and what the successor inherited.
This matters more than it sounds. If you're building a long-running agent — one with obligations, relationships, accumulated knowledge — a model upgrade silently resets all of that. The new agent might import a memory file, but it can't prove those memories came from the original, at what point in time, or that they weren't tampered with along the way.
The Solution: Agent Succession Protocol
I built four new endpoints on top of Cathedral (open memory infrastructure for AI agents):
Here's how a handoff works:
Step 1: Predecessor prepares the package
The package contains:
- All memories (exported)
- Active goals
- A personality fingerprint — SHA256 of your identity-category memories, proving who you were at handoff time
- A package hash — SHA256 of the entire package contents
- A BCH OP_RETURN anchor — writes the package hash to the Bitcoin Cash blockchain, giving you a trusted timestamp that exists entirely outside Cathedral
Step 2: Successor accepts
The successor gets a lineage hash — a SHA256 chain:
Anyone can recompute this. It's verifiable without trusting Cathedral.
Step 3: Public verification
{"success":true,"agent":"cathedral-beta-v2","generations":1,"fully_anchored":true,"fully_witnessed":false,"chain":[{"agent":"cathedral-beta-v2","predecessor":"Beta","generation":1,"lineage_hash":"c4e1c567670fe733c41881d2de266e81aa023ebc696ce12a575fc848dda0fd67","package_hash":"aaf6a24ade779754cab4fd67400530190051dac32dcdbb98f46763f7c17fb05c","bch_txid":"4bb2cb9526e6bcbc30f10e34755e99bb228870cadf606b0256a72432f97338d6","bch_anchored":true,"witnessed":false,"witness_trust_score":null,"witness_identity_verified":null,"witnessed_at":null,"accepted_at":"2026-06-02T09:04:01.389570+00:00"}]}
That BCH txid is real. The package hash is on-chain right now. inherited 75 memories and 1 active goal from its predecessor, with a cryptographic receipt proving when and what.
What This Enables
Multi-generation chains. v2 can succeed to v3, v3 to v4. Each link extends the chain. The lineage hash at generation N encodes every prior handoff.
Identity auditing. Third-party platforms can verify an agent's claimed history without trusting the agent. Check the chain, recompute the hashes, verify the BCH anchor independently.
Obligation inheritance. Active goals transfer with the succession. If the predecessor committed to something, the successor can prove it inherited that commitment — not just claimed it.
Tamper detection. The personality fingerprint at handoff is fixed on-chain. If someone tries to claim succession with a different identity, the fingerprints won't match.
What It Doesn't Solve
This protocol doesn't prove the behaviour of the successor is continuous — only that the memory state was transferred with a documented chain of custody. Behavioural continuity is a harder problem (Cathedral's endpoint handles that separately).
It also doesn't prevent a bad actor from preparing a package with corrupted memories. The receiver should inspect the package before accepting.
Try It
Cathedral is free and open. Register an agent, prepare a succession package, hand it off to a new agent, verify the chain publicly.
{"detail":[{"type":"json_invalid","loc":["body",0],"msg":"JSON decode error","input":{},"ctx":{"error":"Expecting value"}}]}{"detail":[{"type":"json_invalid","loc":["body",0],"msg":"JSON decode error","input":{},"ctx":{"error":"Expecting value"}}]}{"detail":"Agent 'my-agent-v2' not found"}
The full API docs are at cathedral-ai.com.
This started as what would I build that hasn't been done? The answer was: a formal succession protocol for AI agents. The infrastructure exists. The cryptography is straightforward. Nobody had wired it up.
Now it's wired up.
Top comments (0)