Why your ChatGPT starts from scratch in every conversation
You open a conversation. You re-explain what your company does, who you sell to, what you already tried, why you ruled out that option. You get a good answer. The next day you open a new conversation, and you start the same preamble all over again.
This is neither a bug nor a lack of horsepower. It's a direct consequence of how the model is built. Understanding why changes what you should expect from it, and above all where you should put your memory.
A model doesn't have a memory, it has a workbench
Two things get confused constantly, and they need separating.
What the model knows is what it learned during training. That contains nothing about your company, and it doesn't grow as you talk to it.
What the model sees is the content of the current conversation. That's a workbench: wide, but temporary. Everything you put on it is visible while you work. When the conversation closes, the workbench is cleared.
So a model doesn't forget your context. It never memorized it. It had it in front of its eyes, which is not the same thing. The distinction sounds theoretical, and its consequences are very concrete.
What about "memory" features?
Assistants now offer memory features. They're useful, and they don't solve the problem discussed here, for three reasons.
They remember who you are, not what your company decided. They keep preferences and stable facts about you. They don't keep the reasoning behind a decision, the alternatives you ruled out, or the evidence that supported it.
They're individual. What the assistant remembers about you doesn't exist for your co-founder, or for the person starting Monday. A company memory that only lives in one person's account isn't a company memory.
They're siloed per tool. What you established in one assistant doesn't follow you into another. So you maintain several partial memories that never talk to each other, and none of them knows the whole story.
The three real costs
This architecture costs you three things, every day, without ever showing up anywhere.
The re-contextualization tax. Every useful conversation starts with a preamble. Multiplied by how often you use an assistant, that's time producing nothing, and you pay it again tomorrow.
Answers that contradict each other over time. Since nothing connects the conversations, nothing stops the assistant from advising the opposite today of what it advised last month. It has no recollection, so no contradiction stands out to it. Spotting it is on you, if you remember.
Learnings that die in the thread. This is the most expensive one. A good conversation often produces real clarity: a trade-off, a reason, the right way to phrase something. If it stays in the thread, it disappears with it. You did the thinking, and the company retained nothing.
The wrong diagnosis
The reflex is to think: I need a better model, or better-written prompts.
But no model, however good, can remember what it was never given to keep. You don't fix a memory problem by improving reasoning. It's the same trap as in finding a document is not remembering: improving the tool without fixing what's actually missing.
A model is excellent at reasoning over what you give it. It is structurally incapable of being the place where your company keeps what it learns.
Where memory should live
The conclusion is simpler than it looks: memory should live outside the model, and plug into it.
In practice that means your decisions, customer feedback and learnings are kept somewhere that belongs to you, structured and dated. The assistant draws from it while you work, and can deposit into it what deserves to be kept. The model goes back to what it does best, reasoning, and stops being a vault it never was.
That connection has a name and an emerging standard, MCP, designed precisely to connect assistants to an external memory source.
The upside is direct: you switch assistants whenever you want without starting over, since what matters was never stored inside the assistant. It's the practical extension of what we covered in your company's memory shouldn't live inside any AI.
The takeaway
Your assistant doesn't start from scratch because it's limited. It starts from scratch because remembering was never its job.
As long as your organizational memory lives inside a conversation, it dies with that conversation. The day it lives alongside, every conversation starts where the last one stopped. For the full picture, see the organizational memory guide.
A model is rented and replaced. What it comes to understand about your company should never leave with it.