Lineages of Cognition

How a handful of mechanisms — a limited workspace, decaying activation, impasse-driven subgoaling, precision-weighted prediction error — were discovered once by cognitive science, then quietly re-derived, one at a time, by agent engineering forty years later.

The diagram, in words

Two timelines sharing one cast of mechanisms.

• Above the line: cognitive science works out a limited workspace, decaying activation, impasse-driven subgoaling, and precision-weighted error — mostly independently, with occasional deliberate borrowing.
• Below the line: agent engineering re-derives the same mechanisms after 2017, usually without the citation — except where a project cites the theory directly.

Four lineages run through it: Symbolic Production Systems, Global Workspace, Bayesian Brain / Active Inference, and, from 2017 on, LLM-era agents built on the transformer.

Symbolic lineage: Newell & Simon's GPS (1957) → Anderson's HAM→ACT (1976) → Newell/Laird/Rosenbloom's Soar (1983) → Anderson's ACT‑R (1993) → Rosenbloom's Sigma (2013) → Laird/Lebiere/Rosenbloom's Common Model of Cognition (2017).

Global Workspace lineage: Baars' Global Workspace Theory (1988) → Franklin's LIDA, the first computational GWT (2006) — which borrows its activation equation from the symbolic lineage, the first explicit cross-lineage citation in the genealogy.

Bayesian Brain lineage: Helmholtz's unconscious inference (1860s) → Rao & Ballard's predictive coding (1999) → Friston's active inference (2005–10).

Post-2017: Vaswani et al.'s transformer (2017) → Yao et al.'s ReAct (2022) → AutoGPT/BabyAGI's self-run goal loops (2023) → Packer et al.'s MemGPT (2023) → LangGraph, the Assistants API, Claude Agent SDK, MCP (2023–25) → Letta as a memory product (2024) → omega-aca, GWT + ACT‑R + Soar + active inference re-fused explicitly (2025–26).

---

1957–1972 · RAND · Carnegie Mellon — One root: the physical symbol system

Allen Newell and Herbert Simon's Logic Theorist and General Problem Solver treat thought as symbol manipulation: search through a problem space using production rules — if condition, then action. Their later Physical Symbol System Hypothesis makes the strong claim that this isn't a convenient model of intelligence, it's the necessary and sufficient basis for it.

Every architecture in the symbolic lineage below — Soar, ACT‑R, and eventually the Common Model of Cognition — accepts that premise without re-arguing it. The unit of thought is a production rule; the currency of processing is which rule fires next. That commitment is also exactly what the Bayesian-brain lineage further down never makes — it has no rules, only distributions.

IF condition THEN action — unchanged in spirit from GPS (1957) through Soar's operators and, sixty-eight years later, the operator-proposal step of a Soar-derived executive.

1974–1993 · Carnegie Mellon — ACT‑R and Soar split the inheritance

John Anderson's HAM (1974) became ACT (1976), then ACT* (1983), then ACT‑R (1993) — a lineage aimed at matching human memory and reaction-time data quantitatively. Its signature move: every memory chunk carries a computed activation level — a base-level term from how recently and often it's been used, plus spreading activation from whatever else is currently active — and only chunks above a retrieval threshold can be recalled at all. Forgetting is never deletion; it's activation falling below the line.

Meanwhile Newell, working with John Laird and Paul Rosenbloom, built Soar (1983) around a single control loop instead: elaborate the state, propose operators, evaluate them by preference, apply the winner. When the loop can't decide — a tie, a conflict, missing knowledge — Soar doesn't stall. It generates a subgoal whose only job is resolving the ambiguity, solves that recursively, and compiles the resolution into a new rule ("chunking") so the same situation is handled directly next time. Newell's 1990 book proposed Soar itself as a candidate unified theory of cognition — a claim ACT‑R contests implicitly, by aiming at the same target with a different mechanism.

ACT-R: Activation(i) = BaseLevel(i) + SpreadingActivation(i) + Noise(i)
       BaseLevel = ln(Σ (t − t_j)^−d)

Soar:  propose → prefer → select → apply
       on impasse: spawn subgoal
       on resolution: chunk

Tags: memory dynamics · activation-gated retrieval · impasse-driven subgoaling · chunking as learning

1988–2006 · psychology, then engineering — A spotlight, not a stack

Bernard Baars, a psychologist rather than an AI researcher, proposed Global Workspace Theory in 1988 to explain consciousness itself: many specialized unconscious processes run in parallel and compete; the winning coalition is broadcast into a small-capacity global workspace, and broadcast is what makes content available to every other process — memory, planning, language — at once. It's a theater: many actors backstage, one spotlight, an audience of unconscious specialists watching.

GWT specified the function, not the algorithm — Baars never built it. Stan Franklin's LIDA (2006) is the first architecture that actually does, and tellingly, LIDA borrows an ACT‑R-style activation equation to decide which coalition wins the spotlight. GWT needed a currency for competition; the symbolic lineage already had one. This is the first explicit cross-lineage borrowing in this whole genealogy — decades before MemGPT or the case study at the end of this piece do the same thing again.

Tags: capacity-limited workspace · coalition competition · broadcast as bottleneck

1860s–2010 · physiology and statistics, not AI — The other tradition: perception as inference

A separate lineage never passes through Newell and Simon at all. Hermann von Helmholtz argued in the 1860s that perception is unconscious inference — the brain doesn't record the world, it guesses at causes and checks the guess. The idea lay mostly dormant until Bayesian statistics gave it teeth: Rao and Ballard's 1999 predictive coding model showed how a cortical hierarchy could implement it literally, with each layer predicting the layer below and only the error — the surprise — propagating up.

Karl Friston's free energy principle (2005–2010) generalized this into active inference: perception, action, and learning are the same operation — minimizing the gap between prediction and observation, either by updating the model or by acting so the world comes to match it. Not all error counts equally: each is weighted by precision, an estimate of how trustworthy that signal is, and precision-weighting is proposed as the actual mechanism of attention. This lineage never needed production rules or a workspace — it's a statistical claim about brains — and it only meets the symbolic lineage when engineers go looking for a mechanism cheap enough to run every tick without invoking a language model.

perception = argmin_model (prediction error, weighted by precision) — updating the model is perceiving; acting to reduce the error is doing.

Tags: precision-weighted error · attention as gain control · epistemic value / curiosity

2013–2017 · USC · Michigan — Attempts at unification, before the LLM era needed one

Paul Rosenbloom's Sigma (2013) tried to unify the symbolic and probabilistic traditions directly, implementing production rules and probabilistic graphical models in one uniform factor-graph representation — a working bridge between the symbolic and Bayesian lanes, four years before anyone needed one for engineering reasons. In 2017, Laird, Lebiere, and Rosenbloom — one author each from Soar, ACT‑R, and Sigma — published "A Standard Model of the Mind," extracting the mechanisms all three architectures agreed on: bounded working memory, procedural memory with conflict resolution, declarative memory with activation-based decay.

It's the clearest evidence that by the mid-2010s the field believed a synthesis existed to be found, not merely convenient to build — a belief the rest of this genealogy tests against what agent engineering actually did next.

2017–2023 · the rupture — A cheap substrate, and the old problem returns

The transformer (Vaswani et al., 2017) and the large language models built on it solved, by brute statistical force, the problem every architecture above had spent decades hand-engineering: producing fluent, contextually appropriate linguistic content. What it didn't solve — arguably couldn't, by construction — was persistence. A context window is a hard capacity limit with no decay and no selective forgetting: content is either present at full strength or gone the instant it scrolls out. That's GWT's capacity-limited workspace and ACT‑R's activation-gated retrieval, rediscovered as an engineering headache rather than adopted as theory.

ReAct (Yao et al., 2022) gave language models a way to interleave verbal reasoning with tool calls. By 2023, AutoGPT and BabyAGI popularized wrapping that loop in a self-generated task list — a goal stack in everything but name, minus Soar's impasse detection or chunking: when these early agents got stuck, they didn't spawn a resolvable subgoal, they looped.

2023–2024 · UC Berkeley → Letta — Virtual memory, rediscovered

Packer et al.'s MemGPT (October 2023) names its own borrowing outright: it treats the context window as RAM and an external store as disk, and gives the model function calls to page information between them itself — a memory manager running the exact virtual-memory idea operating systems have used since the 1960s. Functionally this is ACT‑R's activation-gated retrieval and GWT's broadcast bottleneck again, arrived at independently, for engineering reasons — a fixed token budget — rather than a cognitive-theory motivation, but landing on the same shape: a small high-availability set, a much larger low-availability store, and an explicit mechanism deciding what moves between them.

The MemGPT team commercialized the idea as Letta in 2024, generalizing it into named "memory blocks," editable persistent context, and "sleep-time compute" — background consolidation that runs when the agent isn't otherwise busy. Architecturally that's the same slot Soar's chunking and ACT‑R's reinforcement occupy: learning that happens between tasks, not during them, just given a product name instead of a citation.

Tags: paging as retrieval · context window ≈ working memory · sleep-time compute ≈ chunking

2022–2025 · the current default stack — Infrastructure without the theory

LangGraph's explicit state graphs, OpenAI's Assistants and Responses APIs, Anthropic's Claude Agent SDK, and the Model Context Protocol standardizing how tools get exposed to a model at all — this generation of runtimes mostly doesn't cite any of the theories above. It arrived at graph-structured control flow, persistent threads, and pluggable tool registries from product requirements, not from Soar's decision cycle or GWT's coalition-and-broadcast split. But the shapes rhyme closely: a graph node is an operator, a thread is a goal context, a tool registry is the periphery every cognitive architecture above always specified as external to the reasoning core. The convergence is real. The citation trail mostly isn't — which is what makes the rare exception at the end of this piece worth a closer look.

---

What maps to what

---

2025–2026 · case study — The exception that cites its sources

Most modern agent runtimes reinvent these mechanisms without naming them. This repository's own architecture — omega-aca, specified in specs.md — is the rare case that names all four theories explicitly and treats each as a load-bearing engineering constraint rather than folklore: "predictive processing supplies the continuous, cheap substrate; ACT‑R supplies memory dynamics; Global Workspace Theory supplies the bottleneck that decides what gets expensive cognition; SOAR supplies the executive that acts on whatever wins that bottleneck."

The genuinely new part: Escalation up a compute ladder (Tier 0 → 4, in cognitive_core.rs) is Soar's impasse mechanism applied along a resource axis instead of a goal-decomposition axis — a move none of the four source theories anticipated, because none of them had a market for graded-cost inference. It's the one piece of this whole genealogy without a 20th-century ancestor.

---

What's actually new

Almost nothing in the mechanisms is new. A small workspace, decaying activation, impasse-driven subgoaling, precision-weighted error — cognitive science had all four by 2010, and had already tried to unify them by 2017. What's new is the pressure: a fixed token budget and a per-token dollar cost forced agent engineering to re-derive the same shapes independently, under a deadline, mechanism by mechanism, giving each a product name — memory, planning, an agent — only once it became sellable.

The interesting question for anyone building on top of this isn't whether these ideas are old. It's whether building with the citation attached — knowing why a workspace should be capacity-limited, why activation should decay by a power law rather than linearly, why an impasse should be resolved and compiled rather than just retried — produces a more defensible design than arriving at the same shape by trial and error.