What is Short-term memory (agents)?

Short-term memory is what an AI agent remembers within the session it is in right now. It is the recent conversation, the current goal, and the latest tool results, all held in the model's context window and available for immediate reasoning. Its defining limitation is that it is transient: when the conversation ends or the context window fills and older turns are dropped, that information is gone unless it was written somewhere durable. Short-term memory overlaps heavily with working memory; the distinction people draw is that short-term memory emphasizes recency (what was said lately) while working memory emphasizes active manipulation (what the agent is operating on for the task). Both are bounded by the context window, which is why managing them is a core part of context engineering, deciding what to keep in view and what to summarize or offload. The counterpart is long-term memory: facts from the current session that are worth keeping get written to a persistent store, so they can be retrieved back into short-term context on a future task instead of being lost when the session closes. An agent with only short-term memory starts every new session cold.