What is an Episodic Memory (AI)?

Most language models are stateless between calls: they know general knowledge from training but nothing about what happened in a specific conversation an hour or a week ago. Episodic memory adds a durable log of concrete events ("user asked X on Tuesday and preferred Y outcome") that an agent can query before acting, distinguishing it from semantic memory, which stores generalized facts stripped of the situation they came from. Building it well requires deciding what counts as an event, how to summarize or compress old ones so retrieval stays fast, and how to rank relevance so a query pulls back the right slice of history instead of everything ever logged. The practical payoff is continuity: an agent that remembers a decision, a rejected approach, or a stated preference stops re-asking the same questions and stops repeating mistakes a user already corrected. The failure mode to watch for is unbounded growth — without decay, deduplication, or summarization, episodic stores balloon and retrieval quality degrades even as storage costs climb.