What Is Episodic Memory in AI Agents?

What is episodic memory?

Psychologist Endel Tulving drew the line between episodic and semantic memory in his 1972 chapter "Episodic and Semantic Memory", in Organization of Memory, edited by Tulving and Wayne Donaldson. He wrote that "Episodic memory receives and stores information about temporally dated episodes or events, and temporal-spatial relations among these events," and described semantic memory, by contrast, as "a mental thesaurus." In a footnote he said he knew of no one who had used the term "episodic" memory before.

AI borrowed the distinction through cognitive architectures. In Soar, as summarized in the CoALA framework (Sumers et al., TMLR 2024), "episodic memory stores sequences of the agent's past behaviors." CoALA carries the idea to language agents: "Episodic memory stores experience from earlier decision cycles," which can include "history event flows," "game trajectories from previous episodes" or "other representations of the agent's experiences."

A 2025 position paper by Pink et al., titled "Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents," names five properties of episodic memory that are useful for LLM agents: long-term storage, explicit reasoning, single-shot learning, instance-specific memories and contextualized memories. How episodic memory sits beside the other memory types is covered in Agent Memory: Architecture, Types, and Trade-Offs.

How episodic memory works

Agents write an episode for each meaningful event and read episodes back by time, by subject or by similarity. Generative Agents (Park et al., 2023) kept a "memory stream," a list of memory objects that each hold a natural-language description, a creation timestamp and a most-recent-access timestamp. Retrieval scored each memory on recency, importance and relevance. The same shape fits a table. An illustrative schema:

CREATE EXTENSION IF NOT EXISTS vector;  -- pgvector

CREATE TABLE agent_episodes (
  episode_id  bigserial   PRIMARY KEY,
  agent_id    text        NOT NULL,
  subject_id  text        NOT NULL,   -- the customer, account or ticket involved
  occurred_at timestamptz NOT NULL,
  action      text        NOT NULL,
  outcome     text,
  embedding   vector(1536)
);
CREATE INDEX ON agent_episodes (subject_id, occurred_at DESC);

-- "What happened to this account most recently?" is a range scan, not a similarity search.
SELECT occurred_at, agent_id, action, outcome
FROM agent_episodes
WHERE subject_id = 'acct_1042'
ORDER BY occurred_at DESC
LIMIT 5;

-- "Have we seen a situation like this before?" is a similarity search.
-- $1 is the embedding of the current situation; <=> is cosine distance.
SELECT action, outcome
FROM agent_episodes
WHERE agent_id = 'refund-agent'
ORDER BY embedding <=> $1
LIMIT 5;

Order matters as much as content. A Salesforce Engineering post on the memory system behind its Agentforce agents puts it this way: "Episodic memory is inherently temporal. Agents must preserve event order and understand how outcomes relate across time, otherwise they risk applying incorrect lessons" (Salesforce Engineering). Episodes also feed semantic memory. CoALA describes Reflexion reflecting on failed episodes and storing the result, such as "there is no dishwasher in kitchen," as semantic knowledge for later attempts.

Why episodic memory matters in real-time systems

Episodic memory records the past, so staleness is not its usual failure; completeness and order are. When several agents write episodes about the same customer or account, the log becomes shared. A chat agent that issues a refund and a voice agent that reads the account history seconds later must see the same log, or the second agent may repeat the refund. The record of what happened is only useful to the next decision if the next decision can see it.

Episodes are also evidence, not current state. "Refund issued" in the log says an agent took an action; it does not say whether the refund settled or the balance changed afterward. Gate decisions on current state read from the system that owns it, and use episodes for history, audit and lessons. One agent recalling its own past conversations with one user is a retrieval problem; a vector index, or Postgres with pgvector, handles it. The case for keeping memory separate from live state is made in AI agent memory architecture.

FAQ