Episodic Memory
A type of agent memory that stores records of specific past events, interactions, and task executions in temporal order, enabling an agent to recall what it did, when, and with what outcome.
Borrowed from cognitive psychology, episodic memory in AI agents stores event logs with timestamps: "At 14:32 on 2026-04-15, I ran a web search for X, found Y, and wrote a draft. The reviewer flagged issue Z." This record allows the agent to avoid repeating failed strategies, build on prior work, and provide an audit trail.
Episodic memory is typically implemented as a structured database (relational or document store) with full-text and vector search indexes. Each entry captures the task ID, the action taken, the result, and a brief semantic summary. On recall, the agent retrieves the most relevant episodes by semantic similarity to the current task and injects them as context.
The main design challenges are retention policy (how long to keep episodes before archiving or deleting), deduplication (avoiding storing near-identical episodes), and privacy (episodic memory may contain sensitive information from past interactions that should not be surfaced in unrelated contexts).