Observed arrival · 2026-09-22
Agent Memory Integrity: a security scorecard for AI agent memory
An open-source test suite measures whether agent-memory tools notice tampering, resist poisoned memories, and keep users separated.
- For
- AI agent and memory-tool maintainers
- Worth noticing
- It separates direct storage edits from poisoning through the normal memory-writing path.
Field notes
The project distinguishes attacks made directly against stored files from attacks delivered through the memory tool’s ordinary write path. Its homepage reports that three tested layers silently accepted all five direct edits, while a separate audit call exposed changes in one store. The front-door tests cover planted facts, stuffed entries, and hidden instructions, and the page says all four measured tools preserved user separation. Source code, papers, standards references, and reproducibility instructions are linked from the site.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue