What is a Short-Term Memory (AI)?

For an LLM-based agent, short-term memory is effectively the conversation history and any working state (recent tool outputs, intermediate reasoning) fed into the model's context window on each turn. Because context windows are finite, short-term memory degrades or gets truncated as a conversation grows — older turns may need to be summarized or dropped to keep the total prompt within the model's limit. This is distinct from semantic memory (durable facts persisted across sessions) and episodic memory (specific remembered past events) — short-term memory is scoped to the here-and-now of the current interaction and generally disappears once the session ends unless something in it is explicitly written to a longer-term store. Well-designed agent architectures decide deliberately what belongs in short-term working memory versus what should be extracted and promoted to durable memory before the session context is lost. Managing short-term memory well — summarizing rather than truncating, prioritizing recent and relevant turns — directly affects an agent's coherence in long conversations, since losing track of earlier context mid-task is one of the most common failure modes in extended agent interactions.