Designing effective memory for AI agents is crucial, yet often an afterthought. This article introduces a decision-tree approach to classify information categories and assign them to the correct memory type: working, semantic, episodic, or procedural. The strategy emphasizes that different data types require distinct memory architectures, preventing common pitfalls like over-provisioning or incorrect retrieval.

Working memory handles information relevant only to the current turn or session, often managed by conversation buffers. Semantic memory stores stable, reusable facts and domain knowledge, such as user preferences or business rules. Episodic memory captures the value of past events and interaction history, acting like a log. Procedural memory optimizes task execution by retaining learned routines and efficient problem-solving steps.

The decision tree guides developers through five questions to determine information persistence, session survival, fact stability, and task repetition. For instance, information needing to persist beyond a single turn but not a session points to working memory. Stable facts, essential across multiple sessions, fall under semantic memory, while historical events necessitate episodic storage. The approach aims to create layered memory setups that efficiently manage context and improve agent performance.

Implementing the wrong memory type can lead to slower retrieval or stale data. For example, using a vector store for static facts is less efficient than a structured profile. This method ensures that agent memory is deliberately designed, enhancing context engineering by only retrieving information that meaningfully improves responses.