Episodic memory stores discrete events (this conversation, that tool result, yesterday's decision) as retrievable records, usually indexed so the agent can pull back the relevant episode when a similar situation recurs. It's what lets an agent say 'last time you asked this, we tried X and it broke' instead of starting cold every session — distinct from semantic memory, which stores general facts. Example: a coding agent recalls that this repo's tests need a specific env var because it hit that wall two tasks ago.
ELI5
Think of Episodic Memory as a named building block in a big LEGO app. Memory of specific past interactions — what happened, when — that the agent can recall later.
In practice
Use it when you are mapping how the app is structured or explaining a feature to a teammate or agent. In practice, define the owner, input, output, and failure mode before you rely on it.