Memory management | LLM Context Engineering | Lecture 6 скачать в хорошем качестве

Memory management | LLM Context Engineering | Lecture 6

Want to go beyond just watching? Enroll in the Engineer Plan or Industry Professional Plan at https://context-engineering.vizuara.ai These plans give you access to Google Colab notebooks, interactive exercises, private Discord community, Miro boards, a private GitHub repository with all code, and capstone build sessions where you build production-grade AI agents alongside the instructors. Everything is designed so that you can actually implement what you learn, not just watch it. Enroll here: https://context-engineering.vizuara.ai In this session of the AI Context Engineering Bootcamp, Dr. Sreedath Panat dives into one of the most important ideas in building scalable AI systems: memory architectures. So far in the bootcamp, we have seen how context is constructed, retrieved, compressed, and isolated. But there is one critical question that remains: how does an AI system remember anything beyond a single interaction? This lecture answers that by introducing different types of memory and how they are designed in real-world systems. We begin by understanding the difference between short-term memory and long-term memory in AI systems. Short-term memory exists inside the context window and disappears after each call, while long-term memory is stored externally and persists across sessions. The challenge is deciding what information should be remembered, how it should be stored, and when it should be retrieved. The lecture then explores different memory types used in production systems, including: Conversation memory, which tracks dialogue history and user interactions Entity memory, which stores structured facts about users or systems Semantic memory, which captures general knowledge in vector databases Episodic memory, which records past experiences and workflows We also discuss how memory is written, retrieved, and updated over time, and how poor memory design can lead to issues such as stale information, irrelevant recall, and increased latency. A key engineering focus of the session is on memory retrieval strategies. Instead of blindly loading all past information, modern systems selectively retrieve relevant memories based on the current query. This is often implemented using embeddings and vector search, similar to RAG, but with a focus on user-specific or system-specific knowledge. Another important concept covered is memory evaluation and pruning. As memory grows, systems must decide what to keep, what to summarize, and what to delete. Without this, memory systems become bloated and degrade performance. We also explore how memory interacts with other components such as tools, RAG pipelines, and multi-agent systems, and how orchestrators manage memory across different agents without overwhelming the context window. By the end of this lecture, you will understand how to design memory systems that allow AI agents to move from stateless interactions to persistent, evolving intelligence.