In an era where every AI headline is about a larger model or a faster transformer, it's easy to forget that an older branch of AI has been quietly asking a harder question: what would it take to actually build a mind? That branch is cognitive architecture. Pat Langley's 2017 paper is one of the better maps of the territory — though the map has some telling blank spots.
I · The territory
What Is a Cognitive Architecture?
Most AI work today is analytic — pick one slice of intelligence and push the metrics. Cognitive architecture is synthetic: it asks how the pieces fit together into a whole mind.
Langley pins down the field by listing features that recur across architectures like ACT-R, SOAR, and ICARUS: a split between short- and long-term memory, modular symbolic structures, pattern matching to retrieve knowledge, dynamic composition similar to chaining in formal logic, and problem-space search as the core loop. The defining commitment is that these mechanisms stay constant across time and domains — whatever the task, the architecture is the same.
II · Progress
What's Improved
The paper traces real progress: a shift from purely symbolic to hybrid representations that can carry probabilistic weight; static knowledge bases giving way to systems that learn new rules from experience; early attempts to bridge mental processing with sensorimotor behaviour; and the addition of episodic memory so an agent can hold onto its own history. The field has not been standing still — it has just been working on problems most of the discourse doesn't notice.
III · The gaps
Where the Paper Falls Short
It picks a paradigm and doesn't fully justify it. Architectures built on different principles — especially those drawing on neuroscience — get short-changed. Recent work suggests high-level cognition may emerge from the strategies by which brain regions interact, not just the symbolic structures within them. That doesn't fit Langley's lens, so it barely appears.
Metacognition is underweighted. Several alternative models have made more progress on “thinking about thinking” than the architectures Langley centres — not a footnote if the goal is human-level cognition.
Deep learning is essentially absent. Given the capital and momentum pouring into it, a paper on the future of cognitive architecture can't really skip the question of how — or whether — these approaches integrate.
Deep learning has given us systems that act impressively. Cognitive architecture is still trying to give us systems that are something.
IV · The open frontiers
What's Still Missing
The most valuable part of the paper is its catalogue of what remains unsolved: reasoning under partial information, where most architectures handle the full-information case fine but the messy half-known world is another story; concept creation, where current systems learn new rules but struggle to invent new concepts; creativity, where humans routinely solve problems with methods they've never seen before and almost nothing in the field captures this; emotion, which overrides logic constantly yet barely appears in cognitive models; and metacognition — reflection on past reasoning, and a stable framework that shapes behaviour across situations.
V · What I'd add
Build for Development, Not the Finished Product
A child's mind isn't a smaller adult mind — it's a system optimised to become one. Maybe the goal isn't an architecture that does everything an adult human can, but one flexible enough to grow into one — including the ability to redefine its own syntax as its thinking patterns evolve.
We share huge fractions of our DNA with other primates and even mice; macaques and humans share brain regions for processing basic language structure. The mind, like the body, evolved from something simpler. Cognitive architecture has been oddly uninterested in either thread. And a lot of the field's incoherence comes from every architecture being judged against its own assumptions — treating cognitive architectures as a set of demonstrable competencies, independent of the model's internals, would let us actually compare them.
Closing
Langley's paper is a good guide, especially for getting your bearings in a field that doesn't get much press anymore. Its real contribution is naming the hard, unfashionable problems — creativity, emotion, metacognition, partial-information reasoning — that any serious attempt at general intelligence has to solve eventually.
Based on Pat Langley, “Progress and Challenges in Research on Cognitive Architectures” (2017), with commentary drawing on Newell (1990), Luppi et al. (2022), Kotseruba & Tsotsos (2018), and Russell & Norvig (2020).