Question
Why doesn't ChatGPT remember how I think?
Because it saves what you volunteered, and how you think is the part you never volunteer. And because memory is a budget, and the entries a budget drops first are the rare ones, which are the valuable ones.
Reason one: you never told it
Memory features store what you say about yourself. That is not a criticism, it is the design, and it works. Tell it you are a lawyer in Chicago and it will stop asking.
Now try to write down, right now, the standards that separate your good work from your average work. Not your preferences. The lines you hold.
Most people can produce three or four and then stop, and the ones they produce are the obvious ones they could always have handed to an assistant. The rest are not missing because you are inarticulate. They are missing because a standard does not present itself to you as a rule. It presents itself as a reaction. You see the draft, something is wrong, and you fix it. The rule was never in your head as a sentence. It was in your hands as a correction.
Which means you cannot volunteer it, and a system built on volunteering will never have it.
Reason two: the eviction policy deletes exactly the wrong things
This is the part that gets less attention and matters more.
Memory has a size. Anything with a size has to decide what to drop. The sensible policy, and the one any competent system uses, is to keep what is recent and what comes up often, and to let go of what has not been touched in a long time.
Now look at your own standards and sort them by how often they come up.
| Kind of thing you know | How often it comes up | What it is worth |
|---|---|---|
| Format and tone preferences | Constantly | Low. Anyone could supply these |
| Standing rules about your work | Weekly | Moderate |
| The call you make on a rare exception | Twice a year | Very high. It is why people ask you |
| The thing you learned once, expensively | Almost never | Highest. It is the whole reason you are careful |
The two columns run in opposite directions. Value climbs as frequency falls. The expensive lesson from four years ago governs a decision you will face once, and getting it right that once is worth more than every formatting preference combined.
So the feature does not fail on the ordinary case. It succeeds on the ordinary case, which is what makes it feel like it is working, and then it is empty on precisely the day you needed it.
Why a bigger window is not the fix
The common next thought is that this is a capacity problem and that longer context solves it.
Capacity is about how much can be in front of the model at once. This is about what got written down at all. If the standard never left your hands as a sentence, then a window a hundred times larger holds a hundred times more of the material that does not contain it. You have made the haystack bigger and the needle is still not in the barn.
The same reasoning disposes of uploading your whole archive, which people try next. The archive is output. Output is what your judgment produced, not the judgment.
What would have to be different
Three properties, and none of them is more storage.
- It has to catch what you did not volunteer. Which means taking corrections at the moment you make them, since that is the only moment the standard is visible, including to you.
- It has to be exempt from eviction. A standard used twice a year has to survive the eleven months in between at full strength. Nothing that prunes by recency can hold this.
- It has to be revisable on purpose. When you change your mind, the entries that were resting on the old position have to be found and revisited, rather than sitting there quietly still being applied.
That third one is a genuinely hard engineering problem and is the reason this is not simply a feature the memory products will add next quarter. It is a different data structure with a different job.
Frequently asked
Why doesn't ChatGPT remember how I think?
Two reasons. It saves what you volunteered about yourself, and the things that actually determine whether its output is any good are things you never volunteer, because you do not know you hold them until something breaks one. And memory has a size budget, so it has to drop entries, and the entries it drops are the ones that come up least often, which are your rarest and most valuable standards.
Doesn't it remember my preferences?
It remembers stated preferences, which is a real feature and worth using. A stated preference is something like the language you work in or the format you like. A standard is different: it is the line you will not cross, it is usually conditional, and you almost never state it in advance because it only becomes visible when something crosses it.
Would a bigger context window fix this?
No, and this is the most common wrong turn. A bigger window changes how much can be in front of the model at once. It does not change what got written down in the first place. If the standard was never captured, a window a hundred times larger holds a hundred times more of the material that does not contain it.
So the memory feature is useless?
It is genuinely useful for what it is: continuity of session. It stops you re-explaining your situation every morning, and that is worth having. What it cannot deliver is continuity of standard, which is a different thing wearing a similar name.
What would have to be stored instead?
The corrections, with their reasons, held permanently and not subject to eviction, because a standard that only applies twice a year has to survive the eleven months in between. That is the whole design difference. A memory keeps what is active. A judgment record has to keep what is rare.
Where this sits
The full comparison is at Imprinted AI vs AI memory, and the category is defined at what Imprinted AI is.
Related
- Imprinted AI vs AI memory. The longer version, arguing the missing operation rather than the eviction policy.
- What happens when AI guesses your preferences? What fills the space when nothing was captured.
- The Final Twenty. Why the correction you just made had nowhere to go.
- Belief propagation. What has to happen when something you told it stops being true.
Keep going
The eviction argument has an uncomfortable extension: every product you use is optimising for the decisions you make often, and you are paid for the ones you make rarely.
The full case is The A.I. Business Manifesto. About 23,000 words, free to read on the page, no gate in front of it. If you would rather have the short version, the same page will send you the three fixes and the PDF.
Read The A.I. Business Manifesto
Free either way. Reading it costs nothing and asks nothing.
Last updated: 28 July 2026