Charles Becker and Sarah Withers teach how to build LLM agents that work like operating systems, managing memory by fetching and editing external data to work around context limits. You'll code an agent from scratch using the open-source framework Leda and create agents with long-term memory and custom memory types.
I'm excited to introduce ALMS as operating Systems Agent Memory built in partnership with Letter, a company that supports the open source agentic framework called Letter. Your instructors are the founders, Charles Becker and Sarah Withers. If you've taken one of deep learner AI's earlier prompting courses, you may have found that when you prompt an LLM via API, they do not have persistent memories. For example, if you prompt an OM with hello, my name is Andrew and then ask in the following prompt, what is my name? It won't remember. To simulate memory, you have to include the conversational history in the prompt or in the context window of the large language model. As a applications have grown more sophisticated, managing the information that goes into and out of the limited context window has become increasingly challenging. MAMPGPT developed a very clever way to handle AI agentic memories.
Take the chatbot application. The context window will include conversational history, but as it grows it may need to be compressed or summarized to fit. It may also want to track personal or topic specific information and include that in the context window.
RAG applications are another example. They strive to include the most relevant information in the context to answer queries.
And agentic applications are the most demanding. They potentially need different information for each processing step. And so in the paper MEMGPT towards oms as Operating Systems, the authors propose using an OM agent to manage this context window. Two of those authors are your instructors for this course.
The analogy we use in the paper is that of virtual memory from the computer systems literature. Computers use a technique called virtual memory to store much more data than the physical memory of the computer can. When data in virtual memory is accessed that is not in physical memory, the operating system will fetch that data from disk and load it into physical memory. In this course you'll learn how to achieve a similar result with an LLM agent acting as the operating system. The agent will determine which information should be in the context window and fetch it from the appropriate source.
To understand the fundamentals, you will code an LLM agent that can edit its own memory from scratch. Then you'll learn how to use an open source agentic framework, leda. This framework allows you to put the LLM as operating systems ideas from MMGPT into practice. It lets you build and deploy agents that have long term memory. Agents of LYDA are given tools to track context memory usage, edit memory, and move information between the context window and external memory. You'll create agents, generate chat memories, and create your own custom memory types.
When I read the original MEMGPD paper, I thought it was an innovative and important technique for handling memory for the Open Source letter framework makes these ideas easy for you to use. Please learn about them in this course.
