Episodic Memory — AI Glossary

Reference details

Related terms

• Agentic Memory
• Semantic Memory
• Memory (Agent)

How it works

1. Interactions are written as timestamped records of what happened, so the store is append-only by nature rather than corrected in place.
2. Episodes are usually summarised before storage, because raw transcripts are long and retrieve poorly against a short query.
3. Retrieval is by relevance and recency together, since an old episode can match a query better on wording while a recent one is more likely to matter.
4. Because records are events, a contradiction between two is not an error — a change of address is two true episodes, not one wrong one.
5. Current state must therefore be derived from the episodes rather than read from them, which is the step that distinguishes this from a facts store.

Example: Two true records, one wrong answer

January: the user says they are based in Berlin. July: they say they have moved to Lisbon. Both episodes are accurate records of what happened, and both remain correctly in the log. Asked where the user is based, a retriever ranking on similarity alone may return the January episode because its phrasing matches the question better. The log is not corrupt; the query needed the latest episode rather than the best-matching one, which is why recency has to be part of retrieval and not an afterthought.

Commonly confused with

Episodic memory records what happened; semantic memory records what is true. The distinction has teeth at write time: a customer changing address is a new episode appended to the log, and simultaneously an overwrite of a semantic fact. Store it only as an episode and the old address stays equally retrievable, because similarity search has no notion of which one is current.

When to use it

Reach for it when:

• Continuity across sessions requires knowing what was said or done before.
• The history itself matters — what was tried, what was rejected, what changed.
• Current state should be derivable from events rather than overwritten.

Reach for something else when:

• Only the current fact is needed, where a semantic store answers directly.
• Retrieval ignores recency, in which case old episodes will answer current questions.
• Storage grows unboundedly with no summarisation, making retrieval noisier over time.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Episodic Memory — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path