Short-Term Working Memory vs Long-Term Memory · 1/2
Working memory lives in the context window
Short-term, or working, memory is everything currently loaded into the context window for the task at hand: the user's latest request, the agent's recent reasoning, recent tool results, whatever it needs to decide the next action. It is naturally bounded by the context window's size and naturally temporary, once the task ends or the window is trimmed, that information is gone unless something saved it elsewhere first.
Working memory is cheap to use, since it's already sitting in the prompt and costs nothing extra to access beyond the tokens it occupies. But it's also fragile. It only exists for the duration of the current session, and within that session it competes for space with everything else the agent is tracking.