7 Types of Agent Memory: Working, Semantic, Episodic, Procedural, Retrieval, Parametric, Prospective
LLMs are stateless by default—agent memory bridges that gap. A technical breakdown of where each memory type lives, what it stores, and when to implement it.
• Working memory: in-context, conversation state, shallow retention
• Semantic memory: vector embeddings, facts and relationships
• Episodic memory: timestamped events, conversation history
• Procedural memory: learned behaviors, task execution patterns
• Retrieval memory: external databases, RAG (Retrieval-Augmented Generation) backed
• Parametric memory: model weights, knowledge baked during training
• Prospective memory: future intent, scheduled tasks and reminders