CoALA: The Blueprint for "Thinking" AGI
By Hire Bikila
OpenPsi Project, iCog Labs
In the OpenPsi project at iCog Labs, we know that a Large Language Model (LLM) is an engine, but it isn’t a "mind." To bridge the gap between text prediction and true AGI, we need a structure. The Cognitive Architectures for Language Agents (CoALA) framework provides exactly that.
What is CoALA?
Instead of a simple "prompt-in, text-out" loop, CoALA organizes an AI agent into three human-like dimensions:
Memory: Distinguishing between Working Memory (current context) and Long-term Memory (facts, experiences, and skills).
Action Space: Balancing External Actions (interacting with the world) and Internal Actions (thinking, planning, and retrieving memories).
The Decision Loop: A continuous cycle of observing, deliberating, and executing.
Why It Matters for OpenPsi
At OpenPsi, we are building agents with internal motivations and cognitive synergy. CoALA aligns with our mission by moving beyond "chatbots" toward Reasoning Agents that:
Stop Hallucinating: By grounding responses in a structured Semantic Memory.
Learn from History: By using Episodic Memory to recall past successes.
Plan Ahead: By deliberating internally before taking an external step.
The Bottom Line
AGI won't be achieved by just making models bigger. It will be achieved by making them more organized. By combining the reasoning power of LLMs with the architectural discipline of CoALA, we are building the "brains" of the future.