Agentic systems require mechanisms to maintain state, learn from interactions, and access knowledge beyond the immediate input context. Standard Large Language Models operate with a limited context window, $L_{context}$, making persistent memory essential for tasks requiring long-term coherence or referencing vast information stores.

This chapter details the design and implementation of memory systems for LLM agents. We will examine techniques for managing short-term information, such as conversation history buffers and summarization methods. Subsequently, we cover long-term memory solutions, primarily focusing on vector databases for efficient semantic retrieval of relevant knowledge using embeddings like $e = f(text)$. You will learn about advanced retrieval strategies to improve the quality of information fetched and how to integrate structured data sources like knowledge graphs or relational databases. Finally, we will address practical aspects of memory management, including consolidation, summarization, and designing effective read/write interfaces for the agent. Practical exercises involve integrating a vector database to provide an agent with persistent memory capabilities.