Working versus long-term memory

Working memory is the context held only for the current session or task, while long-term memory is the small, reviewed set of durable records you keep on purpose. Sorting which is which keeps temporary detail out of permanent storage.

Working memory is the context held only for the current session or task, while long-term memory is the small, reviewed set of durable records you keep on purpose. The two serve different jobs: working memory carries the thread of one conversation, and long-term memory carries the few facts you will still want months from now. Say you spend a session asking an assistant to plan a trip. The flight times you compared, the dates you ruled out, and the back-and-forth are working memory that can be dropped once the trip is booked; the one record worth keeping is the final itinerary you reviewed. When you tell the assistant what to save, you are promoting a checked fact into long-term memory and letting the rest fall away.

Builder example

If your assistant treats every passing detail as something to remember, its long-term store fills with stale fragments that crowd out the facts you rely on. A support bot that saves each customer's offhand remark, then surfaces a half-year-old complaint as if it were current, has confused working memory with long-term memory. Tell the assistant to keep the session context for the task at hand and write only reviewed, durable records to the store you query later, so retrieval stays accurate.

Common confusion: A long context window is working memory, not long-term memory. The window holds a lot of text for one call and then clears; long-term memory is the curated set of records you deliberately save and can pull back in a future session.