Among all the added modules to the augmented LLM, memory is a key component needed to go from an LLM to an Agent. By themselves, LLMs are forgetful entities; they do not remember past conversations, nor do they have access to all actions they have taken. If you were to locally load up an LLM and ask it to remember your name, it can’t–not without explicitly giving it memory. In contrast, the interactions that you might have with hosted LLMs, like ChatGPT and Claude, are not regular LLMs. Rather, they are LLMs augmented with modules like memory and tools. Figure 4-1 illustrates this forgetfulness well, as it demonstrates the interaction you might have with a regular LLM. As such, LLMs are stateless, and information is not persisted across calls.
Over the years, there has been significant attention to aspects of agents like tool usage, reasoning LLMs, and multi-agent collaboration. ...
Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month, and much more.
