February 2024 Summaries

The evolution of agent architectures in artificial intelligence is characterized by the development of an Agentic Action Space, which progresses through several phases as new types of actions are integrated into agent design. Initially, agents like those described in the SayCan paper were limited to external actions, known as "Grounding," which involved direct interaction with the external world. The ReAct paper introduced "Reasoning" as an internal action, expanding the action space to include both external and internal actions. Subsequent advancements incorporated Long-Term Memory, allowing agents to perform internal actions such as "Retrieval" and "Learning," which involve interacting with memory rather than the external environment. These fundamental actions can be combined into composite actions, such as "Planning," which utilizes both reasoning and retrieval. The introduction of new actions necessitates modifications in the execution flow of agents, as seen in the transition from ReAct agents to Planner agents. Parallel actions, inspired by human behavior, enable agents to perform multiple actions simultaneously, enhancing their decision-making capabilities. MemGPT exemplifies this by supporting parallel function calling and long-term memory, offering a framework for further exploration of memory-related actions.

Agents, powered by large language models (LLMs), represent a new class of AI systems that interact dynamically with their environments by utilizing common-sense reasoning and memory storage to perform tasks. These agents, differing from traditional chatbots, have evolved to access various tools and adapt to novel tasks without pre-determined prompt chains, making them increasingly relevant for future applications. They employ both short-term and long-term memory structures, analogous to computer memory systems, to manage and utilize information effectively. Short-term memory serves as the main context available during runtime, while long-term memory, which includes episodic, semantic, and procedural types, stores information externally for future retrieval and decision-making. The design of an agent's memory system is crucial and varies according to use cases, such as role-play, customer support, or task execution. As agents become more sophisticated, they are expected to become more integral to daily life, with improvements in core reasoning capabilities and refined designs set to make 2024 a pivotal year for their widespread adoption. Future discussions will delve into the detailed aspects of how agents interact with memory, focusing on retrieval and learning as key actions within their operational framework.