Back to Research Papers

CoALA: Cognitive Architectures for Language Agents

View on arXiv2023Memory & Context Management

Authors: Sumers et al.

Abstract

CoALA provides a unified cognitive architecture for language agents, drawing from decades of cognitive science research. It defines four memory systems that work together to enable sophisticated agent behavior: working memory for active tasks, episodic memory for experiences, semantic memory for facts, and procedural memory for skills.

Key Contributions

  • 1.Unified framework integrating cognitive science with LLM agents
  • 2.Four-tier memory taxonomy: Working, Episodic, Semantic, Procedural
  • 3.Action space definition for agent-environment interaction
  • 4.Decision-making procedures connecting perception to action
  • 5.Modular architecture allowing different implementations per tier

Implementation in Awareness

Awareness directly implements CoALA's 4-tier structure:

**Working Memory** → Task-focused ephemeral storage - Current task description - Scratchpad for reasoning - Active context summary - Pending action items

**Episodic Memory** → Experience-based retrieval - Full conversation history - Tool execution results - State snapshots at turn boundaries - Retrieved by recency + importance + relevance

**Semantic Memory** → Fact-based knowledge - Extracted facts from conversations - Entity definitions and relationships - Long-term knowledge storage - Retrieved by semantic similarity

**Procedural Memory** → Skill-based execution - Learned workflows and patterns - Tool usage patterns - Successful solution templates - Invoked by task type matching

Code Example

interface CoALAMemory {
  working: {
    currentTask: string;
    scratchpad: Record<string, unknown>;
    activeContext: string;
  };
  episodic: ConversationMessage[];
  semantic: ExtractedFact[];
  procedural: LearnedSkill[];
}

Usage in Awareness

Core memory tier system design - defines our memory taxonomy

Related Papers