Episodic Memory

System retrieval published

Also known as: Event Memory, Autobiographical Memory

Definition

Memory of specific events, experiences, and interactions—the "what happened" record. In AI agents, episodic memory stores past interactions, actions taken, and their outcomes, enabling the agent to recall and learn from specific episodes. Episodic memories are typically timestamped and retrievable by similarity to current context.

What this is NOT

Not semantic memory (semantic is facts; episodic is events)
Not the full conversation log (episodic memory is selectively stored and retrieved)
Not procedural memory (procedural is how to do things; episodic is what happened)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Storage of past conversation turns, tool calls, and observations that can be retrieved when contextually relevant. Often implemented as timestamped entries in a vector store, retrieved by semantic similarity.

Sources: MemGPT architecture, Generative Agents paper (Park et al., 2023), Cognitive architectures for LLM agents

cognitive-science

The memory system for storing and retrieving personal experiences—events that occurred at a particular time and place. Distinguished from semantic memory (general knowledge) and procedural memory (skills).

Sources: Tulving: Episodic memory (1972), Cognitive psychology literature

Examples

Agent recalls 'In our conversation on Jan 15, you said you prefer Python over JavaScript'
Generative agent retrieving memories of past interactions with another agent
Coding assistant remembering 'Last time we debugged this file, the issue was X'
Personal AI remembering specific events from past conversations

Counterexamples

Things that might seem like Episodic Memory but are not:

General knowledge like 'Python is a programming language' (that's semantic)
Knowing how to write a for loop (that's procedural)
The current conversation context (that's working memory)

Relations

specializes long-term-memory (Episodic memory is a type of long-term memory)
overlapsWith agent-memory (Part of agent memory systems)
inTensionWith semantic-memory (Different type of memory (events vs. facts))

Implementations

Tools and frameworks that implement this concept:

Zep secondary