Y. Khandelwal of Tech4Biz Solutions published an Internet-Draft on 1 November 2026 proposing a laboratory benchmark for detecting changes to AI agents’ stored memory. “A Benchmarking Method for the Integrity of AI Agent Memory at Rest,” revision draft-khandelwal-bpm-agent-memory-integrity-00, is work in progress with intended Informational status, rather than an adopted standard.[1]

The method assumes an adversary can write to storage but holds none of the memory subsystem’s cryptographic keys. Evaluators seed memory through the subsystem’s interface, stop it, apply one storage edit, restart it and retrieve memory through its normal read path. Ten edits cover content changes, deletion, reordering, insertion, replay, metadata changes, whole-store rollback and restoration of records erased through the subsystem’s own interface.[1]

The draft distinguishes refusal to serve affected memory, serving it with an integrity signal, and silent acceptance. Read-time refusal or reporting receives full credit; detection only through a separate audit receives half credit. Controls require normal reloads and genuine writes to work and confirm that each edit actually landed. Inapplicable cases are excluded rather than counted as passes, and the method sets no acceptable-score threshold.[1]

The abstract reports 31 result rows, including 18 claiming tamper evidence, with the new erasure-restoration edit measured on seven. The saved excerpt does not include the illustrative-results section, so those figures remain the author’s report here.[1]

[1]: A Benchmarking Method for the Integrity of AI Agent Memory at Rest, Y. Khandelwal, IETF Internet-Draft, 1 November 2026.

Sources

  1. A Benchmarking Method for the Integrity of AI Agent Memory at Restwww.ietf.org · 1 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

05c2742bae893ed26e7994fde6776e26a0c2bd9294b088e06bd4e9456249c50f