A family of minds, a memory that fills
What people mean by the model
When a worker says the assistant, they usually mean the surface in front of them. Behind that surface sits a family of models, each with different characteristics. The smallest, fastest, cheapest model is well suited to short, repetitive tasks — classification, extraction, routing, brief replies. A middle model is the general-purpose default, sensible for writing, analysis, and most coding work. The largest, slowest, most expensive model is reserved for problems that genuinely benefit from extended reasoning. The naming changes more often than the structure; the structure is the thing to learn.
The context window as short-term memory
Every model has a finite amount of attention available at any moment, often called the context window. Everything the assistant currently knows — the conversation so far, any files attached, any prior turn it has remembered — sits inside that window. When the window fills, earlier material starts to be lost. The visible signature is a drift in the work: the assistant forgets a constraint mentioned an hour earlier, contradicts itself, or starts asking for information already provided. None of this is a fault. It is a system running near its memory ceiling.
Three commands worth knowing
Most surfaces provide three practical levers for managing the window. A clear command, which discards the current conversation and starts fresh, is appropriate when switching to an unrelated task and worth using more often than people do. A compact command, which produces a summary of the conversation and replaces the full history with that summary, buys time inside a long working session. A resume command, which reopens a prior session at the point it ended, is the right move when returning to a piece of work after a break. The three together are usually all a worker needs.
Context rot, and how to notice it
The drift produced as a window fills has a name in engineering circles: context rot. It is recognisable before it is named. The assistant begins to repeat itself, to drop earlier instructions, to give responses that feel competent but slightly wrong. The instinct is to push harder with longer prompts; the right move is the opposite — clear the window and re-establish the relevant context cleanly. A practical heuristic is to watch the indicator some surfaces provide for how full the window is, and to take a deliberate action when it crosses about three quarters.
When was the last time a long session with an assistant felt like it lost the thread? What happened, and how did you respond?