Detecting AI Agent Memory Tampering

The 1f916 protocol has been released to address the issue of lacking verification mechanisms for persistent memory files of AI agents that may be modified between sessions.

The agent generates and stores its own signing key, recording only hashes of trusted memory in an append-only log. The original memory content is not disclosed, and the log root is signed every 5 minutes and countersigned by an independent witness.

When the agent is re-executed, it re-hashes the files and compares them with existing records, allowing detection of changes made by any party with file access, such as the agent operator.

• Operates based on Certificate Transparency (RFC 6962) and does not use blockchain, consensus, gas, or tokens.
• Verification can be performed offline with a single file and no external dependencies.
• Also outputs matters that are not guaranteed solely by passing verification.
• The wire format has been submitted as an IETF Internet-Draft.