Cognitive Architecture
Lovelace textbook · CC BY-SA 4.0 · computationalcognitivescience.github.io/lovelace/home
A cognitive architecture is a fixed computational infrastructure within which knowledge and skills operate. ACT-R and Soar are the two most influential architectures. Both use production rules (if-then rules that fire on working memory contents) as their central mechanism. The question is not which module does what, but how the modules compose into a unified system that perceives, remembers, decides, and acts. 
The Natural Framework decomposes this into six functional roles.
Production systems
A production system has three parts: a working memory that holds the current situation, a set of production rules (condition-action pairs), and a conflict resolution mechanism that selects which rule fires when multiple rules match. The cycle repeats: match rules against working memory, select one, fire it, update working memory.
ACT-R
ACT-R (Adaptive Control of Thought-Rational) models cognition as the interaction of independent modules: a visual module, a declarative memory module, a procedural module (production rules), a goal module, and a motor module. Each module operates in parallel but communicates through buffers. Only one production rule fires per cycle (~50ms). Declarative retrieval is governed by activation: frequently and recently used facts are retrieved faster.
Soar
Soar models cognition as search through a problem space. When no production rule applies (an impasse), Soar creates a subgoal and searches for a resolution. Once resolved, the result is chunked into a new production rule so the impasse never recurs. This is learning by problem-solving: every solved impasse becomes compiled knowledge. But chunking only covers one of Soar's learning mechanisms; a diagnostic analysis reveals gaps in the architecture's semantic and episodic learning cells, and a prescriptive follow-up proposes consolidation mechanisms to fill them.
Notation reference
| Term | Meaning |
|---|---|
| Production rule | IF condition THEN action |
| Working memory | Current state of activated knowledge |
| Activation | Retrieval strength in ACT-R (recency + frequency) |
| Impasse | Soar: no production matches; triggers subgoaling |
| Chunking | Soar: compile solved subgoals into new productions |
Neighbors
- Diagnosis of Soar โ where Soar's learning cells are empty
- Prescription for Soar โ consolidation mechanisms to fill the gaps
- Memory โ memory as a RESTful interface vs consolidation as async scheduler
- Lovelace Ch.4 โ neural networks as an alternative substrate
- Lovelace Ch.5 โ reinforcement learning for action selection
ACT-R- Soar
Translation notes
The Lovelace textbook covers ACT-R and Soar in detail, including their module architectures, learning mechanisms, and how they have been applied to model tasks like arithmetic, driving, and air traffic control. This page focuses on the shared core: production systems as the computational mechanism, and the key differences (ACT-R's activation-based retrieval vs Soar's impasse-driven learning). The textbook also discusses EPIC, CLARION, and the debate about whether a single architecture can capture all of cognition. One structural reason architectures stall is that consolidation has its own recursive consolidation, and missing that recursion creates learning plateaus.