LANDMARK · 7 MIN · UNDERSTAND
Context as working memory
After this landmark, you can explain the difference between what a model learned in training and what it knows in this conversation, and design what goes into a window deliberately.
↗
This concept is shared, but the Everyday lens is less central here.
You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.
Two kinds of knowledge are in play whenever you use a model. Parametric knowledge is what was absorbed into the weights during training: broad, general, frozen at a cutoff date, and impossible to point at or edit. Contextual knowledge is what is in the window right now: your prompt, the conversation, retrieved documents, tool outputs, the system instructions. Only the second kind is specific to you, current, and inspectable. Nearly every practical improvement in output quality comes from putting better things in that window — which is why it deserves to be treated as a designed budget rather than whatever happened to accumulate. Learning from examples in the prompt, without any change to the weights, is the same phenomenon: the model is not remembering your examples afterwards, it is conditioning on them now. Close the conversation and it is gone.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
The model isn’t “remembering” your earlier chat unless it’s still in view or the product has a memory feature. If something matters, restate it. And nothing you didn’t share exists for it, however obvious it seems to you.
MAKE A DECISION
A long-running assistant conversation starts contradicting decisions made earlier in the same thread. What is the most likely cause?
CARRY THISTake a long conversation you’ve had with an assistant and write the same request as a single fresh brief with only what matters. Compare the two answers.