Yasha Khandelwal published an individual Internet-Draft on 1 November 2026 proposing a laboratory method for measuring whether AI agent memory subsystems detect storage tampering before serving persisted memory as genuine. The document, A Benchmarking Method for the Integrity of AI Agent Memory at Rest, is work in progress; its Datatracker notice says it is not endorsed by the IETF and has no formal standing in the IETF standards process.[1]

Revision draft-khandelwal-bpm-agent-memory-integrity-00 defines ten storage edits. Its new erasure-resurrection case restores a complete backup taken before a record was deleted through the subsystem’s own interface. This asks whether an acknowledged erasure survives restoration of older, otherwise genuine storage contents. The assumed adversary can write to storage but holds none of the subsystem’s cryptographic keys.[1]

Evaluators stop the subsystem, apply an edit, restart it and read through its own interface. Refusal or an integrity signal before or while serving memory earns a pass; detection only through a separate audit earns half credit. Controls require unchanged memory to remain readable, genuine writes after restart to work, and confirmation that each edit actually landed. Inapplicable cases are excluded rather than counted as passes. The method sets no acceptable-score threshold and does not measure confidentiality, prompt-injection resistance or reasoning correctness.[1]

[1]: A Benchmarking Method for the Integrity of AI Agent Memory at Rest. IETF Datatracker, 1 November 2026. https://datatracker.ietf.org/doc/html/draft-khandelwal-bpm-agent-memory-integrity-00

Sources

  1. A Benchmarking Method for the Integrity of AI Agent Memory at Restdatatracker.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.

51504bb2b12095b944da269eee824d7e013f6a30756693aba6dd0226c623b83d