In recent years, the development of Large Language Models (LLMs) has been rapid and remarkable. These models have evolved from doing simple math problems and writing short essays to generating professional grade code and even assisting in scientific discoveries.
Yet despite these advances, most LLMs are still used in a surprisingly limited way. Instead of being structured systems capable of interacting with their environment, they are often treated as little more than powerful chatbots.
What if AIs could go further? What if they could interact with the world through tools and maintain memory beyond a single conversation? This article explores a conceptual framework designed to achieve just that, the Cognitive Architecture for Language Agents or CoALA for short.
The CoALA Framework
The CoALA framework allows AI agents to interact with the world and access tools by structuring the AI agent into three components: Memory, Action Space, Decision Making. These components work together to allow agents to observe their environment, reason about information and take actions.
Memory
When it comes to memory, CoALA organizes the agent’s knowledge into four types:
- Working Memory: Working memory refers to the information that is immediately accessible to the agent. It also connects to the other components like the long-term memories and external interfaces.
- Episodic Memory: Episodic memory stores the agent’s past experiences or history. The agent can refer back to those past experiences to make better decisions in the current situations.
- Semantic Memory: Semantic Memory refers to the agent’s knowledge about the world, including facts about its environment and itself.
- Procedural Memory: Procedural memory refers to the agent’s skills and internal rules. In language agents, this includes implicit knowledge stored in the model's weights and explicit knowledge set inside the agent's code.
Action Space
The action space defines what an agent is capable of doing. Generally, the larger the set of possible actions, the more complex and domain-specific the agent will be.
Actions can be Internal, interacting with memory or External, interacting with the environment.
External Action: refers to actions that affect the environment through the use of tools and taking in information provided from the environment into the agent’s memory after converting the information into text format. The environment could be the physical world, text with humans or agents or fully digital environments.
Internal Action: They are mental actions rather than physical ones. They could be Reasoning, which is done through the LLM, Retrieval, which is getting the relevant information from the long-term memory and adding it to working memory so it can be accessed immediately or Learning, which is storing information into the agents relevant memory type. Both Reasoning and Retrieval are involved in planning and reading the memory while Learning involves writing onto the memory.
Decision Making
In CoALA, decision making follows a continuous loop that allows the agent to interact with its environment and adapt over time. This process consists of four stages:
Observe: The agent gathers information from the environment.
Reason: The LLM analyzes the information, retrieves relevant memories and plan possible actions.
Act: The agent selects and executes the most relevant action, which could be either an internal or external action.
Update Memory: After acting, the agent updates its memory with the new information gathered from the experience.
Conclusion
As LLMs continue to evolve, their role is expanding beyond simple chat interfaces. Modern AI systems are increasingly expected to act as agents that can interact with their environment, store knowledge and make decisions.
The CoALA framework provides a structured blueprint for designing such systems by organizing agents around memory, actions and decision making which can help researchers and developers build more capable AI agents.
