Skip to content

66. CoALA (Cognitive Architectures for Language Agents)

Mini-Project: CoALA Taxonomy: Comparing Architectures via CoALA Framework

A structured comparison of three agent architectures (ReAct, Plan-and-Execute, Memory-Augmented) across the CoALA dimensions of memory, action space, and decision-making strategy, printed as a formatted comparison table.

View on GitHub


Description

CoALA is a conceptual framework (Sumers et al., 2023) that provides a unified way to describe and compare language agent architectures. It models every language agent as having three core components: (1) memory (working memory, long-term episodic/semantic/procedural), (2) action space (internal reasoning actions + external tool/environment actions), and (3) decision-making (the process that selects actions based on memory state). CoALA gives us a common vocabulary to analyze and design agents.

CoALA is not itself an implementation but a taxonomy. It provides the lens through which we can classify ReAct agents (simple memory + tool actions), Plan-and-Execute agents (richer decision-making), and memory-augmented agents. It unifies all patterns in this guide under a common conceptual framework.

When to Use

  • When designing a new agent architecture from scratch
  • As an analysis framework for comparing existing agent designs
  • When teaching or documenting agent system design
  • Architecture reviews and design documents

Core Components

flowchart TD
    A[CoALA Agent] --> B[Memory Module]
    A --> C[Action Space]
    A --> D[Decision Making]

    B --> E[Working Memory]
    B --> F[Long-Term: Episodic]
    B --> G[Long-Term: Semantic]
    B --> H[Long-Term: Procedural]

    C --> I[Internal Actions: Reasoning, Retrieval]
    C --> J[External Actions: Tools, APIs]

    D --> K[Propose Actions]
    D --> L[Evaluate Actions]
    D --> M[Select & Execute]

    style A fill:#F44336,color:#fff
    style B fill:#2196F3,color:#fff
    style C fill:#9C27B0,color:#fff
    style D fill:#FF9800,color:#fff