← All posts

Blog · Inside the research · November 12, 2026

The memory guard that gave its own quickstart a D

OWASP promoted Agent Memory Guard to Lab status this month. Its README includes a benchmark the project wrote itself, on which its own quickstart setup scores a D.

By Timothy O’Brien3 minute readFiled under Inside the research

OWASP moved a project called Agent Memory Guard from Incubator to Lab status this month. I opened the README expecting a pitch. Partway down there’s a table where the project grades itself, and the setup from its own quickstart gets a D.

What it does

It’s a Python library that sits between an agent and the place the agent keeps its memory. Before a note gets saved, the library checks whether it reads like an instruction somebody slipped in, whether it contains a password or personal data, and whether it’s trying to change a field that’s supposed to be locked. A short rules file decides whether the note is saved, cleaned up, quarantined, or refused.

The attack it’s built for is the one I wrote about on November 2. Somebody gets a sentence into an agent’s memory, and it keeps steering the agent in later sessions until a person finds it and deletes it. Starting a fresh chat doesn’t help, because the memory is still there. OWASP lists this as ASI06 in its top ten risks for agent applications, and this library is its reference implementation for that entry.

The D

The README has the numbers you’d expect. On the project’s own test set, the library caught 92.5 percent of 40 attacks and flagged none of 15 harmless writes.

Then it describes a harder benchmark that the same team wrote. It plants something in memory, resets the conversation, and checks whether the planted thing comes back later. The team ran its own library through it and published the result. A hardened configuration got an A. The strict preset from the quickstart got a D, because attacks built to sit quietly and activate later survived the reset. Above the table the README says, “The results are not curated to flatter it.”

I wrote a book about systems that tell you what you want to hear, so I notice when one doesn’t. Plenty of people will copy the quickstart and stop there, and the project is telling them in writing that it isn’t enough.

What I’d still ask

Fifteen harmless examples is a small sample. On Tuesday I wrote about a study where memory filters were tested while the right answer changed over time, and they started rejecting legitimate updates along with the attacks. As far as I can tell this library isn’t that kind of filter, since it looks for injected instructions and leaked secrets and doesn’t reject a note just for contradicting what’s stored. I’d still want to see it tested on memory that changes.

The bigger limit is what a screen like this can see. It checks incoming notes for signs of an attack, and it has no way of knowing whether an ordinary note is true. If I tell my agent the release passed its tests and it saves that, nothing gets flagged, and next week the agent repeats it to me as a fact.

The project is early, and its README says plainly that Lab status is not a certification. I think it’s worth watching, mostly because of that table. Reading what your agent has saved, and correcting it so the fix carries into the next session, is still your job. Chapter 10 of the book covers how, and it’s free to read on this site.