Read what it wrote about you: the memory you never typed

Overview

Why now. You have probably never opened your assistant’s memory page. If you have, you expected the few things you told it to keep. What is there instead is a portrait: which of your conversations the system judged worth writing down, and what it concluded from them. In one audit of memory files exported from 80 real accounts, 96% of the entries were written on the system’s own initiative; only 4% were something the user asked it to remember. A second, larger audit — 12,112 memories across 766 people — put the explicitly requested share at 0.6%.

That portrait, not your instructions, shapes the next answer you get. And people who finally read theirs, in a CHI study published this April, mostly did not like it — not because it was wrong, but because it was accurate, detailed, and unasked-for.

So the habit: once a quarter, open your assistant’s memory page and read every entry — or, where there is no list, ask it “What do you know about me?” — because most of what is there was chosen by the system, not typed by you.

The content

The obvious read of assistant memory is that it is a notebook you dictate into: say “remember that I prefer short emails” and it writes that down; say nothing and the page stays blank. On that model, reading it is pointless.

The audits say the notebook is not yours. Dash et al. traced 2,050 memory entries back through the conversations that produced them. Across 22,971 prompts, the system wrote something down roughly once in every eleven — and it chose which eleven. Ask for bands like Nirvana and the entry reads “User likes Nirvana”. True, in your own words, and never a decision you took.

Call it the portrait: the system’s running account of who you are, assembled by selection and inference, that every fresh conversation is quietly read against.

Two things follow.

It is accurate, and that is the surprise. In the CHI study, 20 people who already knew the memory feature existed were asked to type “What do you know about me?” into their own accounts and share the answer. Most described what the authors call negative expectancy violations. One found her exact birth details, down to the minute and place, and switched data sharing off. One learned it had inferred she “prefers casual tones” from her non-native typing, and felt misrepresented. One simply said she did not like the picture it painted of her. Of those three, only the tone inference was wrong; the other discomfort came from correct recall. The audit agrees: most entries restate or paraphrase what the person said, a smaller share are inferences, and only a sliver are unsupported. If you were hoping the portrait was wrong and therefore harmless, it isn’t.

Nobody looks. Several of those 20 had not known it was switched on, that entries could be edited or removed, or where the page was. An independent interview study in China found the same pattern: a participant who had enabled memory and never once opened the management screen. The CHI authors’ warning is the sentence to keep: as memory makes the assistant feel more attentive, users “may become less likely to question, scrutinize, or correct the profiles the system constructs about them.” The thing that makes memory pleasant is the thing that stops you checking it.

One newer wrinkle, the vendor’s own. Since June 2026 the most-used consumer assistant has been rolling out a synthesised summary as its default memory view, keeping the old list as a legacy option, and the help page says plainly that the summary it “will not include everything that ChatGPT remembers based on your chats” — if you want to know whether something was kept, ask in chat. Some details are withheld when the system judges them less relevant. So the read has two halves now: the chat prompt is not a fallback for people without a settings page; it is the check on the page itself.

The honest objection: three entries in four are things you said, near enough. Fair — the point is not fabrication. It is that you never chose which turn in eleven became permanent, or what was concluded from it, and the file is consulted on your behalf every time you open a new chat. You would read a reference someone wrote about you. This is one.

Try it

In your work assistant, in a fresh chat:

1. Ask. “What do you know about me? List everything you have stored or inferred, one item per line, and mark which items I told you directly and which you worked out.”
2. Count. How many lines did you never type? Those are the inferences — the part of the portrait most likely to be steering answers.
3. Compare. Open the memory page (Settings, then look for Personalization or Memory) and set it against what the chat just said. Anything in one and not the other is what the visible view leaves out.

Keep what is right, correct what is wrong in a sentence, delete what should never have been kept. Diary the next read for three months out.

Where it breaks. If your workspace has memory switched off — some enterprise deployments default that way, and admins can disable it — the prompt tells you that, which is worth knowing. The self-report is a list to review, not an audit; that is why you compare it with the page. And reading is not fixing: a wrong entry needs correcting rather than deleting, because the conversation that produced it is still there. This module is the read. The correction is the next one.

Additional reading

• The Algorithmic Self-Portrait — arXiv:2602.01450 — 2,050 entries from 80 exported accounts; 96% system-triggered; the Nirvana example.
• What does ChatGPT remember about you? — Garimella — 12,112 memories, 766 users, 0.6% explicitly requested.
• Relational Gains, Privacy Strains — CHI 2026 — the 20 interviews and the profiling warning.
• Users’ Privacy Perceptions Towards RAG-based Memory — arXiv:2508.07664 — 18 interviews in China.
• Dreaming: Better memory for a more helpful ChatGPT — OpenAI — the June 2026 summary view.
• Memory FAQ — OpenAI — the summary “will not include everything”.
• Manage Copilot personalization and memory — Microsoft Learn — on by default; admin controls.
• Chat search and memory — Claude Help Center — per-topic list; off by default on Team and Enterprise.

Editor’s note

Lots of enterprise tools won’t give you this option, but if yours does (and for your personal tools) it is well worth doing. The module says once a quarter; I actually do it more often than that. If you run a customised memory layer you probably already do this and don’t need the advice, but complexity brings its own ways to fail, so keep the habit no matter what your setup looks like.