Cognitive architectures: Soar, ACT-R and blackboard systems

A cognitive architecture is a fixed computational framework (the memories, representations and processing cycle that stay the same across tasks) intended to model the structure of the human mind and to produce general intelligent behaviour once it is given knowledge about a task. Soar and ACT-R are the best-known symbolic examples.

1. What a cognitive architecture is

Allen Newell framed the idea with an equation-like slogan: behaviour is architecture plus knowledge. The architecture is the part that does not change from task to task, like a computer’s instruction set; the knowledge is the program. A cognitive architecture therefore specifies:

• Memories: a short-term working memory for the current situation, and long-term memories, usually procedural (rules for what to do) and declarative (facts and past experiences).
• Representations: symbols, relational structures such as attribute–value graphs or chunks, and sometimes numeric quantities attached to them.
• A processing cycle: how the system matches knowledge against the situation, chooses one action, and applies it.
• Learning mechanisms: how experience adds to or tunes long-term memory.
• Interfaces to perception and action, often with timing parameters taken from human data.

Two motives drive the field and pull in different directions. Cognitive modelling wants an architecture whose behaviour, including its errors and response times, matches people; ACT-R and EPIC lean this way. General AI wants an architecture that can do anything a person can; Soar leans this way. Reviews of the field sort architectures into symbolic, emergent (neural) and hybrid [20]. This page covers the symbolic tradition and the hybrids that grew out of it. The rule-matching machinery they share with expert systems is explained on expert systems.

2. Roots: production systems as a theory of mind

2.1 Production-system cognition

What it is. The claim that human thinking can be modelled as a production system: a working memory holding the current contents of attention, a long-term store of condition–action rules, and a cycle that fires one matching rule at a time. Allen Newell and Herbert Simon developed it in Human Problem Solving (1972) by fitting production-system models to think-aloud protocols of people solving puzzles, cryptarithmetic and chess problems [2]. It extended their earlier General Problem Solver, which used means–ends analysis (see AI search algorithms).

Why productions. They fit what protocols showed: small steps, each triggered by what is currently in mind; a limited working memory; and knowledge that is modular, so a new rule can be added without rewriting the rest. Limits. A production-system model can fit almost any single task by choosing its rules, which makes one model weak evidence for a theory. That problem motivated the next step.

2.2 Unified theories of cognition

In a 1973 paper titled “You can’t play 20 questions with nature and win”, Newell argued that psychology’s habit of testing one binary distinction at a time would never add up to an account of the mind [3]. His answer, set out in his 1987 William James Lectures and the book Unified Theories of Cognition (1990), was to build one architecture and hold it responsible for many tasks at once, from reaction times to problem solving to learning [1]. He offered Soar as a candidate. The demand that one fixed architecture explain everything is what separates a cognitive architecture from a collection of task models.

3. Blackboard systems

3.1 The blackboard architecture

What it is. A blackboard system solves a problem with many independent specialists, called knowledge sources, that never call each other. They communicate only by reading and writing a shared, structured workspace, the blackboard, which holds the current partial solutions organised into levels of abstraction. A control component watches the blackboard, decides which knowledge source to run next, and so steers the search [5]. The metaphor is a group of experts around a blackboard, each adding a piece when they see something they can contribute to.

How it works. Each knowledge source has a condition (a pattern on the blackboard) and an action (changes to the blackboard). When a change triggers several knowledge sources, their activations wait on an agenda, and the scheduler rates them, for example by how reliable the evidence is and how close it is to the current focus. Reasoning is opportunistic: it can go bottom-up from data or top-down from expectations, wherever the best-supported hypothesis is at the moment. That makes blackboards suited to problems where data are noisy, many kinds of knowledge are needed, and no fixed order of steps works.

Limits. The control problem, deciding which specialist to run, is itself hard, and most blackboard systems solved it with hand-tuned scheduling heuristics. Performance and correctness depend on those heuristics, which are hard to validate.

3.2 Hearsay-II

Hearsay-II, developed at Carnegie Mellon between 1971 and 1976 within DARPA’s five-year Speech Understanding Research program, was the first blackboard system [4]. Its blackboard held competing hypotheses at a hierarchy of levels, from acoustic segments through syllables and words to phrases, each tagged with a time interval of the utterance and a credibility rating. Separate knowledge sources proposed syllables from segments, words from syllables, and word sequences from words, predicted likely neighbouring words from grammar, and verified or rejected them. The system could start from an “island” of confidently recognised words and extend it in both directions. Lee Erman, Frederick Hayes-Roth, Victor Lesser and Raj Reddy’s 1980 account presents it both as a speech system and as a general framework for coordinating independent processes. Its lasting contribution was the architecture more than its recognition accuracy; statistical methods soon dominated speech recognition.

The architecture spread quickly. HASP, led by H. Penny Nii at Stanford, applied it to interpreting ocean sonar signals, and Nii’s two-part 1986 survey in AI Magazine became the standard reference [5].

3.3 BB1 and blackboard control

Barbara Hayes-Roth’s BB1 (1985) attacked the control problem by giving control its own blackboard [6]. Control knowledge sources write control plans (strategies, current focus, scheduling policies) to a separate blackboard, and the scheduler follows whatever plan is currently posted. The same opportunistic reasoning that solves the domain problem therefore also decides how to solve it, and the system can explain its strategy. BB1 was applied to construction planning, protein-structure analysis and patient monitoring. The blackboard idea also survives outside AI as one of the standard software architectural patterns.

4. Soar

4.1 Soar

What it is. Soar began as John Laird’s 1983 thesis at Carnegie Mellon with Allen Newell and Paul Rosenbloom, and the standard early description is their 1987 paper in Artificial Intelligence [7]. Its central hypothesis is the problem space: all goal-directed behaviour can be cast as search through a space of states by applying operators. Soar is developed today at the University of Michigan under Laird, written in C and C++, and free under a BSD licence.

How it works. Working memory is a graph of symbolic attribute–value elements rooted in the current state. Long-term procedural knowledge is production rules, matched in parallel with a Rete-style matcher. The decision cycle repeats:

1. Input and elaboration: perception updates working memory, and every matching rule fires in parallel to elaborate the state and propose operators, attaching preferences to them (acceptable, reject, better, worse, best, worst, indifferent).
2. Decision: a fixed procedure reads the preferences and selects exactly one operator.
3. Application: rules that test the selected operator fire and change the state; output commands go to the environment.

Unlike an expert-system engine, Soar has no conflict-resolution strategy that picks a rule. Rules fire in parallel and only propose; the one deliberate choice per cycle is of an operator, and it is made from explicit preferences. Since the 2000s Soar has added semantic memory, episodic memory, reinforcement learning of operator preferences, and a spatial-visual system, described in Laird’s 2012 book [11].

4.2 Impasses and universal subgoaling

When the preferences do not determine a single operator, Soar reaches an impasse: a tie (several acceptable operators, nothing to choose between them), a conflict (contradictory preferences), a no-change (no operator proposed, or the selected one cannot be applied), or a constraint failure. The architecture responds the same way every time: it creates a substate whose goal is to resolve the impasse, and problem solving continues there with all of Soar’s knowledge available, for example by looking ahead to evaluate the tied operators. Because any impasse at any level can spawn a substate, Laird, Newell and Rosenbloom called this universal subgoaling [7]. Planning, reflection and hierarchical task decomposition all arise from this one mechanism rather than being separate modules.

4.3 Chunking

What it is. Chunking is Soar’s original learning mechanism: whenever processing in a substate produces a result for the state above, Soar compiles that processing into a new production rule, a chunk, so that next time the result is produced directly, without an impasse [8]. The idea grew from Newell and Rosenbloom’s 1981 proposal that chunking explains the power law of practice, the finding that the time to perform a task falls as a power function of the number of times it has been practised [9].

How it works. Let $r$ be a result that a substate adds to the superstate. Soar traces back through the rules that fired in the substate to find the set $B$ of superstate working-memory elements the result depended on. It then builds the rule

where the bar marks variablization: identifiers specific to this episode are replaced by variables, so the chunk applies to any situation with the same relevant features. Elements the substate examined but did not use are left out, so the chunk generalises over them.

Worked example. In a blocks-world task, operators move A onto B and move A onto the table tie. In the substate, Soar simulates each, finds that only the first reaches the goal, and returns the preference “move A onto B is better”. Backtracing finds that the result depended on three facts: the goal has A on B, B is clear, and A is clear. The chunk is: if the goal has x on y, and x and y are clear, prefer moving x onto y. The next time a similar tie arises, the chunk fires during elaboration, the preference is there before the decision, and no impasse occurs.

Relation to other learning. Rosenbloom and Laird showed in 1986 that chunking implements explanation-based generalisation: the backtrace is the explanation, and variablization is the generalisation [10]. (Explanation-based learning is covered on symbolic machine learning.) Recent versions of Soar describe their mechanism as explanation-based chunking [12]. Limits. Chunks can be overgeneral when the substate relied on knowledge the backtrace does not capture, and early Soar systems suffered from the utility problem: so many chunks that matching them cost more than the search they saved. Chunking also only compiles what the system could already work out; it does not add new knowledge about the world.

5. ACT-R

5.1 ACT-R

What it is. ACT-R (Adaptive Control of Thought–Rational) is John R. Anderson’s architecture at Carnegie Mellon, the latest in a line that runs from the HAM model of memory (with Gordon Bower, 1973) through ACT (1976) and ACT* (The Architecture of Cognition, 1983) [13] to ACT-R, introduced in Rules of the Mind (1993) [14] and developed with Christian Lebiere and others; the current major version, ACT-R 7, dates from 2015. Its founding premise is that knowledge comes in two irreducible kinds: declarative (facts, stored as chunks: small typed records such as addition-fact: 3 + 4 = 7) and procedural (production rules).

How it works. Specialised modules (visual, manual, declarative memory, goal, imaginal) each communicate through a buffer that holds one chunk at a time. A central production system matches rules against the contents of the buffers and fires one rule per cycle, a cycle of about 50 ms in the standard setting: the serial bottleneck of cognition. Rules request actions from modules, for example “retrieve a chunk matching 3 + 4 = ?”, and modules work in parallel. In the 2004 “integrated theory of the mind”, Anderson and colleagues mapped modules and buffers to brain regions and used the model to predict fMRI activity [15]. What distinguishes ACT-R from a pure symbolic system is its subsymbolic layer: numeric quantities attached to symbols that determine which chunk is retrieved, how fast, and which rule is chosen.

5.2 Activation and base-level learning

Each chunk $i$ has an activation $A_i$ that determines whether and how fast it can be retrieved [24]:

Here $B_i$ is base-level activation: $t_k$ is the time since the $k$-th use of the chunk and $d$ is a decay parameter, conventionally 0.5. The sum $\sum _j{W}_j{S}_{ji}$ is spreading activation from the elements $j$ of the current context, weighted by attention $W_j$ and association strength $S_{ji}$; $\epsilon$ is noise; $T_i$ is retrieval latency with scale $F$; and $\tau$ is the retrieval threshold.

Worked example. A fact used 10 s ago and 100 s ago has $B_i=\ln ({10}^{-0.5}+{100}^{-0.5})=\ln (0.316+0.100)\approx -0.88$. Each further use adds a term, so practice raises activation; each term shrinks as a power of elapsed time, so memories fade as a power law rather than exponentially, matching classic forgetting data. Because latency falls exponentially with activation, practised facts are recalled faster. These equations are why ACT-R models make quantitative predictions of accuracy and response time. Procedural choice is subsymbolic too: each rule has a utility $U_i$, learned from rewards by $U_i\leftarrow U_i+\alpha (R_i-U_i)$, and the matching rule with the highest noisy utility fires.

5.3 Cognitive tutors

ACT-R’s best-known application is tutoring. Anderson’s group wrote production-rule models of how students solve problems in Lisp programming, geometry and algebra, then built tutors that trace each student’s steps against the model (model tracing) and give feedback when a step matches a known misconception or no correct rule. In their 1995 review of ten years of development, the best evaluations showed students reaching the same proficiency as with conventional instruction in about a third of the time [16]. The algebra tutors were commercialised and used in schools across the United States. They are an unusually clear case of a symbolic model paying off: the rules are the curriculum’s skills, written down.

6. EPIC, CLARION and the Common Model of Cognition

6.1 EPIC

EPIC (Executive Process–Interactive Control), by David Kieras and David Meyer at the University of Michigan, couples a production-rule cognitive processor, which unlike ACT-R’s can fire many rules in parallel, with detailed perceptual and motor processors whose timing comes from human-performance data [17]. It was built to predict how long people take on multi-task and human–computer interaction tasks, continuing the engineering-model tradition of Card, Moran and Newell’s GOMS [23]. EPIC’s perceptual-motor modules influenced ACT-R’s.

6.2 CLARION

CLARION (Connectionist Learning with Adaptive Rule Induction On-line), from Ron Sun’s group, is a hybrid. Each of its subsystems has two levels: a top level of explicit symbolic rules and chunks, and a bottom level of implicit knowledge in neural networks [18]. Its signature is bottom-up learning: skills are first learned implicitly at the bottom level, and explicit rules are then extracted from them, modelling how people can perform a skill before they can state it. CLARION is a useful counterpoint to Soar and ACT-R, which start from explicit rules.

6.3 The Common Model of Cognition

In 2017 John Laird, Christian Lebiere and Paul Rosenbloom proposed a standard model of the mind, now usually called the Common Model of Cognition: the consensus that ACT-R, Soar and Rosenbloom’s Sigma had converged on after decades of separate development [19]. Its commitments: a working memory that acts as the interface between components; a procedural long-term memory of rules, with a cognitive cycle of roughly 50 ms that selects one action at a time; a declarative long-term memory; perception and motor modules; and learning that is incremental and happens as a side effect of performance. It is a statement of agreement, not a new system, and a target for comparison with neuroscience and robotics.

7. Comparison

8. Timeline

9. What cognitive architectures are used for today

• Modelling human behaviour. ACT-R and related architectures remain standard tools in cognitive psychology for building models that predict accuracy, response times and learning curves, and in cognitive neuroscience for predicting brain activity.
• Human–computer interaction and human factors. Architecture-based models estimate how long an interface or a cockpit task will take, and where attention and workload will limit people, before anyone is tested.
• Education. Model-tracing tutors descended from ACT-R’s cognitive tutors.
• Simulation and training. TacAir-Soar flew simulated fixed-wing missions in DARPA’s Synthetic Theater of War exercise in 1997, a 48-hour exercise that was then the largest fielding of synthetic agents in a joint battlespace [21]. Soar-based systems are still used for simulated adversaries and teammates.
• Robotics and interactive systems that need explicit goals, memory of past episodes and explanations of their choices.
• Language-model systems. Designers of LLM-based agents have borrowed the vocabulary directly: working memory, episodic and semantic memory, procedural skills, and a decision cycle. The CoALA framework (2023) maps these systems onto cognitive-architecture concepts [22]. The resemblance is architectural; the language model replaces the rule-based decision procedure, and with it the guarantees that came from explicit rules.

10. Limits

• No general intelligence yet. Four decades of work produced impressive integrated systems, but none approaches human breadth. Newell’s unified theory remains a programme, not a result.
• Knowledge still has to be written. An architecture is empty without task knowledge, and most of that knowledge is hand-coded rules. The knowledge-acquisition bottleneck applies here too.
• Grounding and perception. Symbolic architectures assume perception has already produced symbols; connecting them to raw sensors is left to other components.
• Identifiability. A flexible architecture plus freely chosen rules can fit many data sets, so a good fit is weaker evidence than it looks, the problem Newell’s unified-theory demand was meant to fix.
• Scaling learned knowledge. Chunking and similar mechanisms can accumulate rules whose matching cost outweighs their benefit.

11. Cognitive architectures and fail-safe models

A fail-safe model is an AI model whose failures drive it to a controlled, safe state: it abstains when evidence is missing, and learning can narrow what it does but never widen what it is authorised to do. Cognitive architectures offer two relevant ideas. The first is the impasse: Soar treats “my knowledge does not determine what to do” as an explicit, detectable state with a defined response, rather than letting the system pick something anyway. A fail-safe model needs the same honesty, with a different response: where Soar searches for more knowledge, a fail-safe model abstains when its evidence runs out. The second is the separation between a fixed architecture and changeable knowledge, which makes it possible to state properties that hold whatever is learned.

They also show the risk. Chunking adds rules automatically, and a chunk is authoritative the moment it is built, even when it is overgeneral. A fail-safe design puts a gate between learning and authority: a model may propose; only the floor admits a fact. Peel, by Perslis Research, is built on that principle and is, to our knowledge, the first fail-safe model; the exact claim and the closest earlier work are on What is a fail-safe model? There is no neural network in the loop that decides, knowledge is typed, sourced cards, and learning is readable counts. Peel is a research prototype and makes no claim to be a cognitive architecture or a model of the mind. For the paper that argues why chains of steps need a symbolic layer, see The Orchestration Gap; for the broader field, see What is symbolic AI? and the history of symbolic AI.

12. Questions

What is a cognitive architecture?

A cognitive architecture is a fixed computational framework, made of memories, representations and a processing cycle that stay the same across tasks, intended to model the structure of the mind. Given knowledge about a task, it should produce intelligent behaviour on that task. Soar and ACT-R are the best-known symbolic examples.

What is the difference between Soar and ACT-R?

Soar aims mainly at general intelligent behaviour: parallel rules propose operators, one operator is chosen per cycle from preferences, and impasses trigger subgoals and chunking. ACT-R aims mainly at modelling human data: one rule fires per cycle, and numeric activation and utility values predict which facts are recalled, how fast, and which rules are chosen.

What is chunking in Soar?

Chunking is Soar's learning mechanism. When Soar resolves an impasse in a substate and returns a result, it traces which facts the result depended on and compiles them into a new rule, called a chunk. Next time the same situation arises, the chunk fires directly and no impasse occurs.

What is a blackboard system?

A blackboard system solves a problem with independent specialist modules called knowledge sources that communicate only through a shared workspace, the blackboard, which holds partial solutions at several levels of abstraction. A control component decides which knowledge source runs next. Hearsay-II, a speech-understanding system built at Carnegie Mellon from 1971 to 1976, was the first.

What is ACT-R used for?

ACT-R is used mainly to build models of human cognition that predict accuracy, response times and learning, and to relate them to brain activity. Its best-known application is cognitive tutors for mathematics and programming, which trace a student's steps against a production-rule model of the skill.

Are cognitive architectures symbolic AI?

The classic ones are. Soar, ACT-R, EPIC and blackboard systems represent knowledge as symbolic rules and structures. Many add numeric quantities, such as ACT-R's activations or Soar's reinforcement-learned preferences, and hybrids such as CLARION combine symbolic rules with neural networks.

Are cognitive architectures still used today?

Yes. They are used in cognitive psychology and neuroscience to model human behaviour, in human factors to predict task performance, in tutoring systems, and in military simulation. Their concepts of working memory, episodic memory and a decision cycle are also borrowed by designers of language-model agents.

What is the Common Model of Cognition?

It is a consensus model proposed in 2017 by John Laird, Christian Lebiere and Paul Rosenbloom, originally called the standard model of the mind. It summarises what ACT-R, Soar and Sigma agree on: a working memory, a procedural memory of rules with a cycle of about 50 ms, a declarative memory, perception and motor modules, and incremental learning.

13. References

1. A. Newell. Unified Theories of Cognition. Harvard University Press, 1990.
2. A. Newell, H. A. Simon. Human Problem Solving. Prentice-Hall, 1972.
3. A. Newell. You Can’t Play 20 Questions with Nature and Win. In W. G. Chase (ed.), Visual Information Processing. Academic Press, 1973.
4. L. D. Erman, F. Hayes-Roth, V. R. Lesser, D. R. Reddy. The Hearsay-II Speech-Understanding System: Integrating Knowledge to Resolve Uncertainty. ACM Computing Surveys 12(2):213–253, 1980. doi:10.1145/356810.356816
5. H. P. Nii. Blackboard Systems. Part One: The Blackboard Model of Problem Solving and the Evolution of Blackboard Architectures, AI Magazine 7(2); Part Two: Blackboard Application Systems and a Knowledge Engineering Perspective, AI Magazine 7(3). 1986.
6. B. Hayes-Roth. A Blackboard Architecture for Control. Artificial Intelligence 26(3):251–321, 1985.
7. J. E. Laird, A. Newell, P. S. Rosenbloom. SOAR: An Architecture for General Intelligence. Artificial Intelligence 33(1):1–64, 1987.
8. J. E. Laird, P. S. Rosenbloom, A. Newell. Chunking in Soar: The Anatomy of a General Learning Mechanism. Machine Learning 1(1):11–46, 1986. doi:10.1007/BF00116249
9. A. Newell, P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. In J. R. Anderson (ed.), Cognitive Skills and Their Acquisition. Erlbaum, 1981.
10. P. S. Rosenbloom, J. E. Laird. Mapping Explanation-Based Generalization onto Soar. Proceedings of AAAI-86, 1986.
11. J. E. Laird. The Soar Cognitive Architecture. MIT Press, 2012.
12. J. E. Laird. Introduction to the Soar Cognitive Architecture. arXiv:2205.03854, 2022.
13. J. R. Anderson. The Architecture of Cognition. Harvard University Press, 1983.
14. J. R. Anderson. Rules of the Mind. Lawrence Erlbaum, 1993.
15. J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, Y. Qin. An Integrated Theory of the Mind. Psychological Review 111(4):1036–1060, 2004.
16. J. R. Anderson, A. T. Corbett, K. R. Koedinger, R. Pelletier. Cognitive Tutors: Lessons Learned. Journal of the Learning Sciences 4(2):167–207, 1995.
17. D. E. Kieras, D. E. Meyer. An Overview of the EPIC Architecture for Cognition and Performance with Application to Human-Computer Interaction. Human-Computer Interaction 12(4):391–438, 1997.
18. R. Sun. Duality of the Mind: A Bottom-up Approach Toward Cognition. Lawrence Erlbaum, 2002.
19. J. E. Laird, C. Lebiere, P. S. Rosenbloom. A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. AI Magazine 38(4), 2017. doi:10.1609/aimag.v38i4.2744
20. I. Kotseruba, J. K. Tsotsos. 40 Years of Cognitive Architectures: Core Cognitive Abilities and Practical Applications. Artificial Intelligence Review 53:17–94, 2020. arXiv:1610.08602
21. R. M. Jones, J. E. Laird, P. E. Nielsen, K. J. Coulter, P. Kenny, F. V. Koss. Automated Intelligent Pilots for Combat Flight Simulation. AI Magazine 20(1), 1999.
22. T. R. Sumers, S. Yao, K. Narasimhan, T. L. Griffiths. Cognitive Architectures for Language Agents. arXiv:2309.02427, 2023.
23. S. K. Card, T. P. Moran, A. Newell. The Psychology of Human-Computer Interaction. Lawrence Erlbaum, 1983.
24. J. R. Anderson, C. Lebiere. The Atomic Components of Thought. Lawrence Erlbaum, 1998.