What is agent memory and how do different memory types work?

Agent memory enables AI agents to maintain context and learn from interactions. Short-term or working memory holds the current conversation context within the LLM context window, including recent messages and tool results. Long-term memory persists information across sessions using external storage, typically vector databases for semantic retrieval of past interactions and knowledge. Episodic memory records specific past interactions and experiences the agent can reference. Semantic memory stores factual knowledge and learned concepts. Procedural memory captures how to perform specific tasks based on past successful completions. Implementation typically uses conversation buffers for short-term memory, vector stores for long-term semantic memory, and structured databases for factual records. Memory management includes summarization to compress older context, selective retrieval to pull only relevant memories, and forgetting mechanisms to remove outdated information. Effective memory design significantly improves agent consistency, personalization, and task performance across interactions.