Episodic memory in large language model systems refers to information learned or retained after initial training, often from user interactions or session context. It may be temporary within a conversation or persistent across sessions, and it contrasts with semantic memory learned during pretraining.
In LLMs, information learned after training. In contrast, semantic memory is information learned during training. Episodic memory can be temporary (for example, lasting only within the current chatbot session) or more permanent (for example, persisting for each session the user invokes). See also procedural memory.