Episodic memory
The agent's memory of specific past events and sessions — "what happened when" — usually stored as timestamped summaries that can be retrieved by time, topic, or participant.
Episodic memory answers questions like "what did this user ask me last Tuesday?" or "what was the outcome of the last similar request?" It is event-shaped rather than fact-shaped. Each episode has a timestamp, participants, and a summary.
In production agents, episodic memory is usually a vector store of session summaries with metadata (user, channel, time, outcome). The summaries are written by a small LLM at session end and retrieved on the next session when relevant.
Episodic memory is the layer most responsible for an agent feeling "like it knows me." It is also the most expensive layer to get wrong — stale or hallucinated episodes poison future sessions.
Frequently asked
How is episodic memory different from chat history?+
Chat history is verbatim transcripts. Episodic memory is summarized, indexed, and selectively retrieved. You can fit ten thousand episodic summaries in the context where a hundred raw transcripts would not fit.