AI Agents 6 - Memory, Learning, and Adapation

Описание: In this lecture from CSE 491/895: AI Agents at Michigan State University, Dr. Mohammad Ghassemi covers the foundations of memory management, learning, and adaptation in AI agents.

The session is divided into two main parts:

1. Memory Management
Why agents need both short-term memory (context windows) and long-term memory (databases, knowledge graphs, vector stores) to retain context and perform multi-step tasks.
Types of long-term memory: semantic (facts), episodic (experiences), and procedural (rules).
Practical coding examples showing how agents can store, update, and retrieve memory to guide interactions.

2. Learning + Adaptation
How agents adapt by changing their reasoning or actions in response to new data.
Reinforcement learning related techniques such as Proximal Policy Optimization (PPO), explained step by step (prompt completion, reward modeling, policy updates).
Introduction to Direct Preference Optimization (DPO), which refines models using pairs of preferred vs. rejected responses.

By the end, you’ll understand how to build agents that not only recall past interactions but also learn and improve over time, making them more robust and trustworthy.