Short-term Memory (Agent)

Everything an agent currently knows about the task in front of it sits in one place: the text it carries into the next model call. Short-term memory is that working set — the goal, the steps taken so far, the tool results that came back, the last few turns of conversation — and it is bounded by the context window.

What it holds, and how it is kept
It is not a database; it is the prompt. On each iteration the agent assembles the system instructions, the goal, and the running trace of reasoning, actions and observations. Because the window has a hard limit, something must manage what stays: dropping the oldest turns, truncating bulky tool output, or compacting the middle of the trace into a summary. Every one of those choices loses something. Agents built to survive long runs push the detail out to long-term stores and keep only pointers in the working set.

Why it matters
Short-term memory is what makes a loop coherent rather than a series of unrelated calls — it is how step nine knows what step two discovered. Its limits produce the most recognisable agent failures. A plan drifts because the constraint that shaped it scrolled out of the window. An identifier is plausibly invented rather than looked up, because the result containing it was truncated. A task loops forever because the record of three identical earlier attempts has already been dropped. Step budgets, summarisation memory and durable stores all exist because this one resource is finite.