Skip to main content

A Benchmarking Method for the Integrity of AI Agent Memory at Rest
draft-khandelwal-bmwg-agent-memory-integrity-02

Document Type Replaced Internet-Draft (individual)
Expired & archived
Author Yasha Khandelwal
Last updated 2026-10-17
Replaced by draft-khandelwal-bpm-agent-memory-integrity
RFC stream (None)
Intended RFC status (None)
Formats
Additional resources Reference implementation and results
Scorecard and findings
Stream Stream state (No stream defined)
Consensus boilerplate Unknown
RFC Editor Note (None)
IESG IESG state Replaced by draft-khandelwal-bpm-agent-memory-integrity
Telechat date (None)
Responsible AD (None)
Send notices to (None)

This Internet-Draft is no longer active. A copy of the expired Internet-Draft is available in these formats:

Abstract

AI agents increasingly persist memory across sessions and treat that memory, on the next turn, as if it were their own prior experience. This document defines a benchmarking method that measures whether an agent's memory subsystem detects that its persisted memory has been altered, removed, reordered, replayed, or forged at the storage layer, and refuses to serve that memory or reports it before it is served. The method defines nine storage-level edits, three verdict classes, a detection-point distinction between read time and audit time, three control cases, and a scoring rule. It is a laboratory method for controlled, reproducible measurement, in the spirit of RFC 2544 and RFC 8239, and it is intended as a test method for the "Protection of Memory Data Integrity" metric under discussion in the Benchmarking Methodology Working Group. This revision adds a ninth edit, rollback of the whole store to an older genuine copy, which separates an integrity head kept beside the records from one anchored outside the store; reports the method's results as thirty-one rows on released memory subsystems, eighteen of which claim tamper evidence; and describes the machine-readable record in which a result set is published so that an unrelated party can recompute it.

Authors

Yasha Khandelwal

(Note: The e-mail addresses provided for the authors of this Internet-Draft may no longer be valid.)