Understanding CoALA: A Cognitive Architecture for Language Agents (2023)

Short summary

CoALA structures language agents with four interacting memory types—working (current state), episodic (past experiences), semantic (general knowledge), and procedural (execution skills)—each with independent persistence and retrieval rules. This contrasts with current LLM systems that collapse everything into a single context window, preventing consistent learning and execution. The framework suggests agent intelligence emerges from coordinated memory systems rather than raw model scaling.

• •CoALA introduces four independent memory systems for language agents: working (immediate context), episodic (historical experiences), semantic (factual knowledge), and procedural (execution patterns)
• •Each memory type operates with different persistence and retrieval rules, enabling agents to learn from history and execute consistently across sessions
• •Framework positions agent intelligence as emerging from coordinated memory interaction rather than raw model scaling

Generated with AI, which can make mistakes.