Episodic Memory
People remember incidents. An agent that cannot recall 'the outage last Tuesday' feels broken.
A diary of days, not the wiki of truths you distilled from them.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Recall last time
Scroll inside the frame to explore · use + / − to zoom up to 200%
Query episodes by time or similarity, then summarize for the prompt.
Key Takeaways
- 1.Episodic memory stores events: what happened, when, with whom. Each episode has a time, a context, and an outcome.
- 2.Retrieve episodes when the user asks 'last time we…'. Do not treat an episode as a forever fact — extract facts separately.
- 3.Store episodes with timestamps and a short summary. Retrieve by recency or embedding.
- 4.Promote stable conclusions into semantic memory; keep the story as an episode.
Learn elsewhere
- →Semantic Memory
- →Memory Retrieval
Real Example
Scenario
User: 'the refund we discussed Friday.' Episodic hit: Friday 4pm, order 4412, promised a callback. Semantic fact: that order is still open.
What you would do
In Agent Memory, apply Episodic Memory to this scenario: User: 'the refund we discussed Friday. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Episodes have a whenStore summary + timeExtract facts separatelyRecall by recency or meaning
Cheat Sheet
Quick recap
quick ref- •Diary of events
- •When + what + outcome
- •Not a forever fact
- •Promote facts out
Common Mistakes
- ✕Using episodes as the only memory type
- ✕No timestamps on stored chats
- ✕Never extracting a fact from a repeating episode