# What is Episodic Memory?

Also called Experience Memory, Event Memory.

Episodic memory is the record of specific past events an AI system experienced, such as a particular conversation, a task attempt and its outcome, or a tool call and its result. Each entry is tied to a time and a situation, which lets a system recall what happened on a given occasion rather than only what is generally true.

The term is borrowed from cognitive psychology, where episodic memory covers personally experienced events and is contrasted with semantic memory for general knowledge. AI systems adopt the same distinction. An episode records that a particular deployment failed last Tuesday with a specific error, while semantic memory records that deployments generally fail when a migration is missing.

Episodes are typically stored as timestamped records containing the situation, the action taken, the outcome, and sometimes a short reflection. They are retrieved by similarity to the current situation, by recency, or by explicit reference. Because they are concrete, they support learning from experience without changing model weights, which is why agent frameworks lean on them.

Their value depends on outcome labeling. An episode recording only what was attempted teaches little, while one recording whether the attempt succeeded, and why it failed, lets a later run avoid the same path. Some systems periodically distill many episodes into general rules, which converts episodic memory into semantic memory.

Cost and noise grow with volume. A busy agent generates episodes continuously, most of which will never be useful, so retention policies, sampling, and consolidation are necessary. Retrieval by situation similarity also risks recalling superficially similar episodes whose lessons do not transfer, so precision matters more than exhaustive capture.

## Key points

- Timestamped records of specific events and their outcomes.
- Contrasts with semantic memory for general knowledge.
- Outcome labels are what make episodes useful.
- Distilling many episodes produces general rules.
- Needs retention limits to avoid noisy recall.

## In practice

An agent tries to export a report and the request fails because the file exceeds a size limit. It writes an episode recording the request, the error, and the successful workaround of splitting the export by month. Two weeks later, facing a similar export, it retrieves that episode and splits the job up front instead of failing first.

## Related terms

- [Semantic Memory](/en/glossary/semantic-memory)
- [Agent Memory](/en/glossary/agent-memory)
- [Long Term Memory](/en/glossary/long-term-memory)
- [Working Memory](/en/glossary/working-memory)
- [Short Term Memory](/en/glossary/short-term-memory)

[Back to the AI Glossary](/en/glossary)