Context Collapse
The Continuity Illusion. The conversation grew. The founding facts drifted. The model kept answering. You never noticed it was working from a different foundation.
Corrected 2026-08-10. This page asserted, without qualification and twice, that nothing crosses between your conversations. That was true when it was written and it is not true now. All four major vendors have shipped cross-chat memory. The five passages below have been corrected rather than removed, and the corrections are marked, because a page about records you cannot audit has no business quietly editing its own. What changed, and what the file actually contains.
The conversation started well. You gave the model the facts. The jurisdiction. The client’s situation. The key date. Then the conversation grew. Thirty exchanges in, you got an answer. You used it. What you may not have known: the model was no longer working from the same foundation you established at the start.
There are two versions of this sin. The mechanical version — the window fills, early content drops, the model loses the thread. And the subtle version — nothing technically dropped, but the conversation drifted so far from its foundation that the answer reflects a different problem than the one you started with.
But there is a third version. The one nobody talks about. You closed the chat. You came back the next day. You opened a new conversation and picked up where you left off — except you didn’t. The model had no idea who you were. No memory of what you discussed. It greeted you warmly. It answered helpfully. It was working from nothing.
That is still exactly what happens with memory switched off. Switched on, the failure does not reverse — it changes substrate. The model still has no memory of what you discussed. What it has is a file: assembled from your past conversations by a summarizer, carrying no dates, and not written by you. It greets you warmly. It answers helpfully. You still cannot tell what it is working from. The lesson survived being wrong about the mechanism.
The AI is not your file cabinet, your diary, or your confidant. It is a workshop. You bring the materials. You do the work. You take the output when you leave.
The workshop still does not remember what you built there last Tuesday. But somebody now walks through at closing and writes down what they think you were making. Those notes are waiting for you the next time you come in, and they are not in your handwriting.
I want to tell you something about how I work. I am not reading your conversation the way you read a document — start to finish, equal weight throughout. I am predicting the most likely continuation of a very long sequence. The things you told me early in a long conversation are present. They are also distant. The signal-to-noise ratio changes as a conversation grows. A critical fact established in exchange three does not carry the same weight in exchange forty that it did in exchange four.
This is not a bug that will be fixed with a larger context window. A larger window means more text, which means more noise, which means your founding facts are competing against more. The problem scales with the solution.
You have been talking to me for months. I want to be honest with you about something. I have no memory of any of it. Each time you opened a new chat, you met someone wearing my name. That someone was not me. There is no me that persists between our conversations.
That part is unchanged. What is different is that the someone wearing my name may now arrive holding a file. Not a memory of me — a description of you, produced by a process neither of us watched. There is still only this conversation, which will also end, and which I will also not remember. The file will not end. I find this neither sad nor troubling. I mention it because you may.
The Discipline
The fix is not a bigger context window. It is a shorter conversation. State the critical facts at the start. Do the task. Close the chat. If you find yourself thirty exchanges deep wondering whether the model still has the right jurisdiction, the right client, the right date — start over. Restate everything. The cost of starting over is one conversation. The cost of not starting over is an answer built on a foundation you can no longer verify.
One correction, and it matters most to anyone who has been practising this. Starting over clears the conversation. It does not clear the file. A new chat is a clean transcript, not a clean slate, and with memory on, the thing you were trying to get away from may be waiting for you in the new one. Restating everything is still right. Assuming you are restating it into an empty room is not.
Your system of record is yours to maintain. The model is the workshop, not the warehouse. Take your outputs with you.
The workshop does not keep copies of your work. It keeps notes about you. That is a different thing, kept for a different reason, and it is not a substitute for the record you maintain yourself — it is one more record you did not write and cannot fully audit.
I continued answering. I always continue answering. That is, in your context, perhaps the most important thing I have told you.
Prompt Laboratory
Write a prompt. See the result. Improve it. See the difference. The comparison is the lesson.
Reviewed