Episodic memory is a type of AI agent memory that stores specific past events and interactions as a sequence, preserving what happened and when. Borrowed from cognitive science, the term contrasts with semantic memory, which holds general facts independent of when they were learned.
For an AI agent, episodic memory is what lets it remember the history of its interactions — previous conversations with a user, actions it took, and outcomes it observed. This allows continuity across sessions: an assistant can recall that you asked about a topic last week, or that a particular approach failed earlier, and act accordingly rather than starting fresh each time.
Vector databases support episodic memory by storing embeddings of past events alongside timestamps and other metadata. When the agent needs relevant history, it retrieves episodes by semantic similarity to the current situation, often combined with recency filters. Effective episodic memory usually requires more than raw similarity, layering in time, importance, and summarisation so the agent recalls the right moments rather than simply the most similar text.