Soar

Soar is one of the longest-running projects in artificial intelligence: a general cognitive architecture and the rule-based agent language that comes with it. Rather than a general-purpose programming language in the usual sense, Soar is a fixed computational framework for building intelligent agents — systems that must perceive, decide, act, and learn — together with a production-rule notation for encoding the knowledge those agents use. Begun in 1983 and still actively developed today, Soar embodies a specific and ambitious bet: that a small, fixed set of mechanisms, combined with enough knowledge, is sufficient to produce the full range of intelligent behavior, from routine reaction to deliberate planning to learning from experience.

History & Origins

Soar grew directly out of the research tradition of Allen Newell and Herbert Simon at Carnegie Mellon University, whose earlier work on the Logic Theorist and the General Problem Solver (GPS) established the ideas of heuristic search, problem spaces, and the physical symbol system hypothesis. Soar began around 1983 as the PhD work of John Laird, supervised by Newell, and developed together with Paul Rosenbloom. The name originally stood for State, Operator And Result — reflecting Soar’s basic decision cycle — though the project now treats “Soar” as a proper name rather than an acronym.

The three creators each brought a strand to the design. Newell contributed the long-standing goal of a unified theory of cognition — a single architecture that could account for the breadth of human intelligent behavior rather than a collection of task-specific programs. Rosenbloom contributed chunking, originally a model of how humans speed up with practice, which became Soar’s single learning mechanism. Laird tied these together into a working system, and after joining the faculty at the University of Michigan in 1986, made Michigan the long-term home of Soar’s development. Early versions were implemented in Lisp; the architecture was later reimplemented in C (the Soar 6 line, in the early 1990s) and is maintained today in C and C++.

Newell’s death in 1992 did not end the project; if anything, his 1990 book Unified Theories of Cognition, which used Soar as its central worked example, set an agenda that the Michigan group has pursued ever since.

Design Philosophy

Soar is built around a handful of tightly integrated commitments:

• Problem spaces and operators. All tasks are formulated as searching a problem space: the agent is in some state, applies operators to change that state, and works toward a goal. This single formulation is meant to cover everything from tic-tac-toe to controlling an aircraft.
• Knowledge as production rules. An agent’s knowledge is encoded as if-then production rules that fire in parallel against working memory, proposing, evaluating, and applying operators. Rules do not make decisions directly; they contribute preferences, and a fixed decision procedure selects the next operator.
• Impasses and universal subgoaling. When the available knowledge is insufficient to choose or apply an operator, Soar reaches an impasse and automatically creates a substate — a new problem space in which to figure out what to do. This universal subgoaling means deliberation, planning, and meta-reasoning are not special cases but the natural response to not knowing what to do next.
• Learning by chunking. When reasoning in a substate resolves an impasse, Soar chunks: it compiles the reasoning that produced the result into a new production rule, so that next time the same situation is handled immediately without the impasse. This turns effortful, deliberate reasoning into fast, automatic recognition — Soar’s model of practice and skill acquisition.

The overarching aim is generality with a fixed architecture: the mechanisms never change from task to task; only the knowledge does.

Key Features

Later Soar also incorporates an appraisal-based model of emotion, in which the agent’s assessment of a situation modulates learning and behavior.

Evolution

For much of its first two decades, Soar was deliberately minimal: procedural knowledge (rules), working memory, and chunking as the sole learning mechanism. Landmark work like the 1987 paper “Soar: An Architecture for General Intelligence” and Newell’s 1990 Unified Theories of Cognition defended the position that this compact set of mechanisms could scale to human-level breadth. Applications from this era — R1-Soar (a reconstruction of the R1/XCON configuration expert system, around 1985), Robo-Soar (1991, controlling a Puma arm), and above all TacAir-Soar (flying tactical air missions in the 1997 STOW-97 exercise) — showed the architecture running large, knowledge-rich agents in demanding, sometimes real-time, settings.

The Soar 9 line, opening around 2008 with Laird’s work on Extending the Soar Cognitive Architecture, marked the biggest conceptual expansion. Recognizing that chunking alone could not capture every form of human memory and learning, the Michigan group added reinforcement learning, distinct semantic and episodic long-term memories (Derbinsky and Laird, 2009–2010), the Spatial Visual System for imagery, and an appraisal-based emotion model. These additions kept the original symbolic core but surrounded it with the additional memory and learning systems a general agent seemed to require. Laird’s 2012 MIT Press book, The Soar Cognitive Architecture, consolidated this modern design.

Development has continued steadily since, through the Soar 9.6 series, with the most recent release, Soar 9.6.5, appearing in May 2026 — largely an engineering and build-system modernization (a first-class CMake build, multi-platform continuous integration, and language bindings for Python, Java, Tcl, C#, and JavaScript).

Current Relevance

Soar is developed and maintained today by John Laird’s research group at the University of Michigan, with continuing work through the Center for Integrated Cognition. It is distributed as free, open-source software under a BSD license, runs across major desktop platforms, and can be embedded in host applications through its language bindings and integrated with robotics middleware such as ROS. Its natural home is research: cognitive modeling, autonomous agents, robotics, interactive task learning (as in the Rosie agent), and game AI.

Alongside ACT-R, Soar is widely regarded as one of the two most influential general cognitive architectures, and it helped shape the Common Model of Cognition — an effort to identify the points of consensus across cognitive architectures. It is more a research instrument and living theory than a mainstream production tool, but its longevity and continued release cadence are unusual for any software project, let alone one born in 1983.

Why It Matters

Soar is important less as a language to write everyday programs in and more as a sustained argument about the nature of intelligence. It took Newell and Simon’s foundational ideas — symbols, search, problem spaces, production systems — and pushed them toward a complete, running architecture that could learn and act, then held that architecture stable for decades so that it could be tested, extended, and compared. In doing so it pioneered ideas now common in AI and cognitive science: automatic subgoaling in response to lack of knowledge, learning by compiling deliberate reasoning into fast rules, and the integration of multiple memory and learning systems within one agent. Few systems have pursued a single scientific goal — understanding general intelligence through a fixed set of mechanisms — as consistently, or for as long, as Soar.