Note: This article is a summary and interpretation of the research paper
Cognitive Architectures for Language Agents
(2023) by Michael Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Rather than proposing a new architecture, the goal here is to explain the paper's core ideas in an accessible way and explore why they matter for the future of AI memory systems.
Modern language agents feel intelligent, but under the hood they are still fragile systems stitched together with prompts, context windows, and external tools. The CoALA framework (Cognitive Architectures for Language Agents) proposes a more structured view: instead of treating LLMs as standalone reasoners, we should treat them as components inside a cognitive system with memory, actions, and decision loops.
At the center of this framework is a simple idea borrowed from cognitive science: intelligence depends heavily on how memory is structured.
CoALA breaks memory into four interacting systems: working memory, episodic memory, semantic memory, and procedural memory.
Working Memory: The Active Scratchpad
Working memory is the agent’s current “mental workspace.”
It contains everything the model is actively using right now:
In CoALA terms, this is the active state used for the current decision cycle. It is not persistent. Once the cycle ends, it disappears unless explicitly stored elsewhere.
You can think of it as the agent’s RAM fast, limited, and constantly overwritten.
Most LLM systems today stop here. That is why they feel intelligent but inconsistent: they reason well in the moment, but forget the structure that produced good decisions.
Episodic Memory: What Happened Before
Episodic memory stores experiences over time.
Instead of storing abstract facts, it stores:
Example:
The agent tried to deploy a service, forgot database migrations, caused a rollback, then fixed it by adding a pre-deploy migration check.
This is important because intelligence is not just knowing facts it is learning from sequences of events.
In CoALA, episodic memory supports learning from history, enabling agents to improve across repeated tasks instead of resetting every session.
Without episodic memory, agents repeat mistakes endlessly.
Semantic Memory: What the Agent Knows
Semantic memory is the system’s store of general knowledge.
This includes:
Example:
PostgreSQL requires schema migrations before deployment in production systems.
Unlike episodic memory, semantic memory is not tied to a specific experience. It is decontextualized knowledge that can be reused across situations.
In modern systems, this is often implemented through:
Semantic memory is what allows agents to “know things” even if they have never directly experienced them.Procedural Memory: How to Do Things
Procedural memory stores skills and action patterns.
This is not about facts or past events. It is about behavioral structure:
Example:
To deploy safely: run tests → build → apply migrations → deploy → verify health checks.
This type of memory is critical because it encodes execution ability, not just knowledge.
In CoALA, procedural memory can live in:
Without procedural memory, an agent may “know” what to do but fail to consistently execute it.
Why These Four Matter Together
The key insight in CoALA is that intelligence emerges from coordination between memory types, not from scaling a single context window.
Most current LLM systems collapse all of this into one space: the prompt. CoALA argues that this is fundamentally limiting.
Instead, memory should be treated as a structured system with different persistence rules, update mechanisms, and retrieval strategies.
Closing Thought
CoALA reframes language agents not as chat models, but as cognitive systems with modular memory and structured decision loops. It connects modern LLM tooling back to decades of cognitive science, showing that we are rebuilding ideas that were already explored in symbolic AI just now with language models as the core engine.
This raises an interesting question: if AI can organize memories like humans, can memory do more than just retrie
Reference:
Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. Cognitive Architectures for Language Agents. arXiv, 2023. https://arxiv.org/abs/2309.02427