The assistant
It should rememberwhy you are long,not every wordyou typed
Context windows forget quietly and at the worst moment. Memory here is explicit: written cards you can read, correct and delete, carried into the turns where they matter.
Why explicit beats implicit
A long conversation degrades before it fails
Push enough into a context window and the model does not announce that it has lost the thread — it answers slightly worse, referring to a plan it half-remembers. Explicit cards mean the thesis is re-supplied rather than recalled, and you can see exactly what was supplied.
The right thing to remember is small
Not the transcript. The thesis, the invalidation, the constraint you keep repeating, the thing that went wrong last time. A card is a fact worth carrying into future turns, and the discipline of writing one is what keeps memory from becoming a second, worse transcript.
You can read and correct it
Memory you cannot inspect is a system that will confidently act on a misunderstanding you never saw it form. Cards are visible, editable and deletable, and a wrong one is a bug you can fix in a sentence rather than a mood the assistant is now in.
A watch reads the same memory
When a scheduled re-check runs, it lands in the session it came from with the cards attached. That is what lets a re-check say "this invalidates what you said on Tuesday" rather than "the price moved".
Questions about memory
Is my memory shared with anyone?
No. Cards belong to your account, and the platform refuses a request for an account it cannot confirm is yours.
Does it remember across sessions?
Cards do — that is their purpose. A session is a conversation; a card is a fact that outlives it.
What if it remembers something wrong?
Delete or edit the card. That is a deliberate advantage of explicit memory: the failure is visible and one-line fixable, rather than a slow drift you diagnose by re-reading a month of chat.
Does memory cost tokens?
Yes, and that is a reason to keep cards short. A memory system that grows without bound makes every future turn more expensive — the cost of context is the constraint that keeps this honest.