Episodic Memory

Definition

Episodic memory in AI agents mirrors the human cognitive concept: a persistent, indexed record of specific past events that can be recalled to inform current behavior. Unlike working memory (the current context window) or semantic memory (general knowledge), episodic memory captures what happened in prior sessions with a specific user or on a specific task. It provides agents with the ability to remember previous conversations, learn from past errors, and build relationships across time.

Engineering Context

Episodic memory allows agents to learn from past interactions. Implementation: store interaction summaries (compressed to 200-300 tokens each) with metadata and embeddings. When a new task arrives, retrieve the top-k most semantically similar past episodes and inject their summaries into the working context. Requires TTL policies to prevent indefinite growth and namespace isolation between users. For multi-tenant systems, episode namespacing by user_id is essential—episode bleed between users constitutes a privacy violation. Common storage patterns: PostgreSQL with pgvector for smaller deployments, dedicated vector databases (Pinecone, Qdrant) for scale. Episode compression using summarization models prevents token budget exhaustion over time.

Related Terms