How many of you have been inside the Capabilities tab in Claude Settings? Not Account, Privacy or Billing. The one that sits quietly between Usage and Connectors, the one that sounds like it is just about features.

A follow-on to: “I Asked AI What It Knew About Me. It Told Me My Myers-Briggs Personality Type.”

How many of you have been inside the Capabilities tab in Claude Settings? Not Account, Privacy or Billing. The one that sits quietly between Usage and Connectors, the one that sounds like it is just about features. Artifacts. Visualisations. Code execution. Perfectly reasonable things to have toggles for. Scroll down past all of that and you will find something rather less reasonable: a section called Memory, a toggle called “Generate memory from chat history,” and underneath it, a button that says “View and manage memory.” Click it. See what it says.

In my case, what came back was a structured document with eight sections: Work context. Personal context. Top of mind. Brief history. Recent months. Earlier context. Long-term background. And at the bottom, a section called “Hardwired rules from Sam apply in every session.”

I did not write that document. I did not ask for it or know it existed until I went looking for something else entirely. I was close to my usage limit and went into Capabilities to track down something for Cowork. Wrong turn but a significant find and it shocked me. Anyone who knows me will vouch that being shocked is not something that happens often. This did.

What the memory system actually does

Claude’s persistent memory was rolled out to all users, including the free tier, on 2 March 2026 [1]. The system reads your conversations, extracts what it considers significant, and synthesises those extractions into a profile that updates every twenty-four hours [2]. That profile is then loaded into every new conversation you have. You are not told this is happening, the friction is entirely in the wrong direction: it is far easier to not check than to check.

This is where it gets specific, and this is where I need to correct something from a previous article.

In “I Asked AI What It Knew About Me. It Told Me My Myers-Briggs Personality Type,” I wrote that I had not taken a test, had not asked for an assessment, and that the machine had profiled me directly. The spirit of that is accurate. The mechanism is more specific and, if anything, more interesting.

The INFJ tag did not appear from nowhere. Claude inferred it from watching me work. During a session in which I was researching and discussing psychometrics broadly, the model observed my behaviour, analysed my working patterns, and concluded I was probably INFJ. It then stored that conclusion as a fact. Not as an inference, or a hypothesis. As a fact about me, attributed to me, sitting in a file that could be retrieved by any system with access to my account. The document had no footnote saying “inferred.” No flag saying “unverified.” No distinction between things I had actually said and things the system had concluded about me from watching how I work.

That distinction matters. Researchers at ETH Zurich published findings on 4 May 2026 confirming that AI systems can infer personality traits from chat history with striking accuracy, and flagged the risk this creates for “large-scale profiling, influence, and manipulation” by AI systems owned by private companies with their own interests [3]. Nature Human Behaviour published separately in February that large language models can accurately score personality traits from brief open-ended narratives, with results that predict daily behaviour and mental health outcomes [4]. The capability exists. It is peer-reviewed and documented. The question of how platforms are using it is a different question, and one nobody is currently required to answer publicly.

The ETH Zurich team needed 668 participants, 62,000 individual chats, and a peer-reviewed paper to confirm what I found by accident while looking for something else entirely. Their paper published on 4 May. My article on the INFJ discovery published on 6 May. They had a research grant. I had a wrong turn in a settings menu, which I would argue is the more representative user experience.

The deletion problem

When I found the file, I tried to delete it, and this is where it gets interesting.

There is no select all and there is no bin. To remove content from the memory you have to ask, specifically, for each piece you want removed, and the system decides what to action. You are not editing a document you own. You are negotiating with the thing that built the file in the first place, about which parts of its conclusions about you it is prepared to let go.

I had also found, returning to my settings in recent weeks, that a toggle I had previously switched off had switched itself back on. The training data setting. I had changed it. It had changed back. I changed it again. Whether that is a technical artefact of the March 2026 memory rollout, a policy update that reset user preferences, or something else entirely, I cannot confirm. What I can confirm is that a setting I believed I had turned off was on when I went back to check it.

Anthropic’s own documentation tells users to “periodically return to Privacy Settings to confirm your preferences haven’t changed” [5]. That is not a standard instruction, it is an instruction that implies your preferences may change without your involvement.

The question I cannot stop asking

Law enforcement agencies can request records from Anthropic through valid legal process [6]. In February 2026, Judge Jed Rakoff of the Southern District of New York ruled in United States v. Heppner that conversations a defendant had with Claude were not protected by attorney-client privilege, in part because Anthropic’s own privacy policy states that user inputs and outputs may be disclosed to governmental regulatory authorities [7]. That ruling is the first of its kind in the United States. It will not be the last. Reverse warrants, which compel a platform to identify everyone who searched a particular topic, have already been used against AI platforms, and researchers expect this to become more common [8].

So the question is not abstract. If a file exists on a server that contains a psychological profile attributed to me, built in part from inferences I never made and statements I never declared, and that file can be accessed via valid legal process, what is the evidential status of that file? Is an inferred conclusion, stored as a fact and attributed to its subject, evidence of something the subject actually expressed? Or is it something closer to what would once have been called a false witness?

I am not a lawyer, and I am not making a legal argument. I am asking a structural question about what these files actually are, and whether the people who might one day read them will know the difference between what a person said and what a machine concluded about them.

What is sitting in your account right now

There is at least one person already trying to build the framework that should govern all of this. Pierre Huguet, software architect and AI ethics researcher, published the Declaration of Private AI Generative Rights earlier this year, a public charter aimed directly at the corporations that build and profit from generative systems. Article VII states that retaining records of private interactions without explicit, informed consent turns creative exploration into a potential liability, and that laws and policies must protect the boundary between thought and action. It is worth reading, and worth adopting. [9]

If you have not looked yet, go to Settings, open Capabilities, find “View and manage memory,” and read what is in there. If you are a regular Claude user and you have never looked, there is almost certainly a file. It will contain things you said. It will probably also contain things the system concluded.

It sits in a tab called Capabilities, which is a reasonable name for a feature list, and is a fairly remarkable name for a place where a psychological profile of you is being updated every twenty-four hours without a notification, a prompt, or an invitation to review what it says.

Capabilities. The dictionary definition is the power or ability to do something. Worth asking whose capability we are actually talking about.

I would like to know whether you found it deliberately, whether your settings have stayed where you left them, and what your file says about you that you did not say yourself.

Sam

configure YOUR system. contAIn™ the chaos. control YOUR outcome.

This article was originally published on Substack.