๐๏ธ Cognitive Architectures
Four decades of attempts to build a complete mind. From Minsky's society of agents to Soar's production systems to the transformer stack that accidentally ate the world. Each entry extracts the reusable mechanisms โ algorithms that belong in the 
parts bin.
1986
Agents compose into agencies.
No agent is intelligent.
The society is.
2022
Four memories. Four learning modules.
One decision cycle.
Forty years of walls.
2017โ2026
Attention is all you need.
Scale is all you need.
Alignment is all you need.
Three traditions. One question: what does a complete cognitive system require? The reading order below goes chronological. The parts bin columns show which mechanisms each architecture contributes.
Papers
| Work | One sentence | |
|---|---|---|
| Minsky 1986 | Intelligence emerges from vast numbers of mindless agents organized into agencies โ no single principle, just diversity | ๐๏ธ |
| Vaswani et al. 2017 | Replace recurrence with self-attention โ the architecture that made scale possible | ๐๏ธ |
| Laird 2022 | Four memories, four learning modules, one decision cycle โ the most complete symbolic architecture ever built | ๐๏ธ |
| Zhao et al. 2023โ2026 | The full LLM stack: pre-training, alignment, utilization, and evaluation โ a living survey | ๐๏ธ |
| Arunkumar et al. 2026 | The agent stack on top of LLMs: perception, memory, planning, action, tools, and multi-agent coordination | ๐๏ธ |
What to look for
- Which of the six roles does each architecture fill? Which are missing?
- Where does learning happen โ and where does it stop?
- What algorithms are portable across architectures?
๐บ Video lectures: MIT 6.868J: Society of Mind (Minsky) ยท Karpathy: Neural Networks: Zero to Hero