Cognitive Architecture Survey

Overview

A cognitive architecture is a systematic description of the internal information processing structure of an agent, defining how it perceives, reasons, learns, and acts. From the early explorations of the 1950s to today's LLM-driven architectures, cognitive architectures have always been a central topic in agent research.

What Is a Cognitive Architecture?

A cognitive architecture is a theory about the fundamental computational structure of intelligent systems. It addresses:

Representation: How does the agent represent knowledge?
Reasoning: How does the agent use knowledge to make decisions?
Learning: How does the agent acquire new knowledge from experience?
Memory: How does the agent store and retrieve information?
Execution: How does the agent translate decisions into actions?

Formally, a cognitive architecture can be defined as a tuple:

\[ \mathcal{A} = \langle \mathcal{M}, \mathcal{R}, \mathcal{L}, \mathcal{P}, \mathcal{E} \rangle \]

where:

\(\mathcal{M}\): Memory system (short-term memory + long-term memory)
\(\mathcal{R}\): Reasoning mechanism (rule matching, search, neural reasoning)
\(\mathcal{L}\): Learning mechanism (chunking, reinforcement learning, gradient descent)
\(\mathcal{P}\): Perception module (input processing and encoding)
\(\mathcal{E}\): Execution module (action selection and output generation)

Three Major Paradigms

graph TD
    A[Cognitive Architecture Paradigms] --> B[Symbolicism<br/>Symbolic]
    A --> C[Connectionism<br/>Connectionist]
    A --> D[Hybrid Architecture<br/>Hybrid]

    B --> B1[SOAR]
    B --> B2[ACT-R]
    B --> B3[BDI/PRS]

    C --> C1[Neural Networks]
    C --> C2[Deep Learning]
    C --> C3[Transformer/LLM]

    D --> D1[ACT-R 6.0<br/>Symbolic + Statistical]
    D --> D2[CoALA<br/>LLM + Structured Memory]
    D --> D3[RAISE<br/>LLM + Cognitive Modules]

1. Symbolic Architecture

Core Idea: Intelligence is symbol manipulation. Knowledge is represented as explicit symbols (rules, logical formulae, semantic networks), and reasoning is achieved through symbol transformation.

Representative Architectures:

Architecture	Developer	Core Mechanism
SOAR	Laird, Newell, Rosenbloom	Problem-space search + Chunking
ACT-R	Anderson	Production rules + Activation spreading
BDI/PRS	Bratman, Georgeff	Belief-Desire-Intention reasoning
Icarus	Langley	Concept hierarchy + Skill execution

Pros: Strong interpretability, rigorous reasoning, editable knowledge

Cons: Knowledge acquisition bottleneck, lack of robustness, difficulty handling perceptual data

2. Connectionist Architecture

Core Idea: Intelligence emerges from the connections of many simple units. Knowledge is distributed across connection weights.

Representative Architectures:

Architecture	Core Mechanism
Perceptron/MLP	Feedforward network + Backpropagation
RNN/LSTM	Recurrent connections + Gating mechanisms
Transformer	Self-attention + Positional encoding
LLM (GPT/Claude)	Large-scale pre-training + RLHF

Pros: Automatic learning from data, strong robustness, handles high-dimensional input

Cons: Poor interpretability, unreliable reasoning, difficult to edit knowledge

3. Hybrid Architecture

Core Idea: Combines the strengths of symbolic and connectionist approaches, using neural networks for perception and learning, and symbolic systems for reasoning and planning.

Modern Hybrid Architectures:

CoALA (Cognitive Architectures for Language Agents): LLM as reasoning core + structured memory modules
RAISE: LLM + explicit reflection, planning, and memory modules
LLM + Tool Calling: LLM's implicit reasoning + precise computation from external tools

Core Components of Cognitive Architectures

Memory System

┌──────────────────────────────────────┐
│         Long-Term Memory (LTM)        │
│  ┌──────────┐ ┌──────────┐ ┌───────┐│
│  │Declarative│ │Procedural│ │Episodic││
│  │ (Facts)   │ │ (Skills) │ │(Exper.)││
│  └──────────┘ └──────────┘ └───────┘│
├──────────────────────────────────────┤
│   Working Memory / Short-Term (STM)   │
│   Current context, active goals,      │
│   temporary information               │
├──────────────────────────────────────┤
│         Perceptual Buffer             │
│   Temporary storage of raw input data │
└──────────────────────────────────────┘

Cross-Reference

For a detailed discussion of memory systems, see Memory Systems Survey.

Reasoning and Decision-Making

The reasoning mechanisms differ significantly across architectures:

Architecture Type	Reasoning Mechanism	Decision Speed	Reasoning Quality
Rule matching	Condition-action rules	Fast	Limited by rule quality
Search	State-space search	Slow	Optimal but computationally intensive
Probabilistic reasoning	Bayesian networks	Medium	Handles uncertainty
Neural reasoning	Feedforward/Autoregressive	Fast	Pattern matching, not logical
LLM reasoning	CoT/ReAct	Medium	Flexible but unreliable

Learning Mechanisms

Architecture	Learning Method	Description
SOAR	Chunking	Compiles successful search paths into direct rules
ACT-R	Activation adjustment	Adjusts memory activation values based on usage frequency and recency
Neural Networks	Gradient descent	Optimizes connection weights through loss functions
LLM Agent	In-context Learning	Learns within the context window without weight updates
LLM Agent	Experience accumulation	Stores reflections in external memory for future retrieval

Evolution from Classic to Modern

graph LR
    A[Classic Symbolic<br/>1960-1990] -->|Knowledge acquisition bottleneck| B[Statistical Learning<br/>1990-2010]
    B -->|Deep learning revolution| C[Neural Architecture<br/>2010-2020]
    C -->|Scaling + Pre-training| D[LLM Agent Architecture<br/>2020-]
    D -->|Structured + Controllable| E[Hybrid Architecture<br/>CoALA/RAISE]

Key Turning Points:

1990s: The knowledge acquisition bottleneck in symbolic systems spurred the rise of statistical methods
2012: AlexNet demonstrated the power of deep learning, fully reviving connectionism
2017: Transformer unified NLP architectures
2022: ChatGPT showcased the potential of LLMs as general cognitive engines
2024: Frameworks like CoALA attempted to understand and improve LLM agents using cognitive architecture theory

Chapter Contents Guide

File	Topic	Core Question
BDI Model	Belief-Desire-Intention	How to formalize rational agent behavior?
ACT-R and SOAR	Classic cognitive architectures	How to unify memory, learning, and reasoning?
LLM Cognitive Architecture	Modern architectures	How do LLMs map to cognitive functions?
World Models and Internal Representations	Internal representations	How do agents simulate and predict the world?
Architecture Design Patterns	Engineering patterns	How to organize the computational components of an agent?

References

Newell, A. (1990). Unified Theories of Cognition. Harvard University Press.
Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.
Laird, J.E. (2012). The Soar Cognitive Architecture. MIT Press.
Sumers, T. et al. (2024). Cognitive Architectures for Language Agents. arXiv:2309.02427.
Kotseruba, I. & Tsotsos, J.K. (2020). 40 Years of Cognitive Architectures: Core Cognitive Abilities and Practical Applications. AI Review, 53, 17-94.