Workshop: Cognitive Architecture Transfer Scan

Question

What concepts, distinctions, mechanisms, and research methods from cognitive architectures might improve Commonplace's account of agent memory, context engineering, learning, orchestration, and knowledge-base operation?

The immediate purpose is breadth-first conjecture generation. The architecture names come from the Wikipedia comparison of cognitive architectures that triggered the discussion. Every listed item gets its own scan, including entries that may turn out not to be cognitive architectures at the same level of analysis.

Epistemic boundary

This pass is deliberately memory-first and source-ungrounded. It records what appears worth investigating before paying the cost of reading original papers, manuals, or code. Statements under "Remembered model" are recollections, not established descriptions. Transfer candidates are conjectures, not recommendations.

Each scan carries a recall-confidence label:

  • High — the main components and mechanisms are stable enough in memory to generate fairly specific questions, but details still require checking.
  • Medium — the broad orientation is remembered, while terminology or component boundaries may be wrong.
  • Low — only a theme, reputation, or name is remembered; the document mainly records questions and guards against invented detail.

Confidence concerns fidelity to the source architecture, not the merit of the transfer idea. A useful conjecture can survive after its alleged precedent is corrected, but it must then be presented as Commonplace's own argument rather than attributed to the architecture.

No claim from this workshop should be promoted with an external attribution until it has been checked against a primary source. Grounding should be selective: first identify a claim that would materially change a Commonplace design or evaluation, then read the most direct paper or manual for that claim.

Comparison boundary

The triggering table mixes at least four levels:

  1. general cognitive architectures;
  2. narrower cognitive models or processing strategies;
  3. implementation frameworks and libraries;
  4. architectures specialized for robotics, affect, teams, or social simulation.

That mismatch is itself useful. Any future comparison should type its subjects before comparing features; otherwise a library can appear to "lack" a theory of cognition and a processing principle can appear to "lack" a software ecosystem merely because unlike things occupy the same rows.

Scan index

Item Recall confidence Strongest provisional transfer
4CAPS medium model capacity per cooperating component and treat overload as redistribution
ACT-R high distinguish stored chunks, cue-driven activation, retrieval, and compiled procedures
ASMO low make attention selection an inspectable arbitration process
CHREST high learn indexes by discriminating cases and treat expertise as retrieval structure
CLARION high govern explicit and implicit competence as interacting but non-identical paths
Copycat high construct representations and analogies through competing micro-hypotheses
Deeplearning4j high type substrate, framework, model, and architecture separately
DUAL medium distinguish the durable graph from the transient coalition that becomes active context
EPAM high turn retrieval failures into new discriminations rather than adding undirected content
EPIC medium-high model tool and reasoning resources separately and study scheduling bottlenecks
GLAIR medium require knowledge-level actions to have an explicit grounding path to execution
Hierarchical Temporal Memory high use violated sequence expectations, not only similarity, to detect novelty
LIDA high separate availability, attentional selection, broadcast, action, and learning
MAMID low-medium represent global control state as threshold and priority modulation rather than content
Mibe low make abstention and source-confidence first-class outputs of broad scans
OpenCog high separate epistemic confidence, attentional priority, utility, and behavioral authority
Parallel terraced scan high allocate deeper reading progressively while preserving exploration
Procedural Reasoning System high separate goals, applicable plans, commitments, and reconsideration conditions
R-CAST medium model teammates' information needs and push only decision-relevant context
Soar high use impasses to open subproblems and compile verified resolutions into fast paths
Sigma medium-high test whether a shared inferential substrate removes real translation costs
SOSIEL medium evaluate methodology across heterogeneous actors and social adoption dynamics
TinyCog low use a minimal executable architecture as an adequacy probe, not a completeness claim

See working synthesis for cross-architecture patterns and the initial grounding queue.

Shared extraction frame

Every item is scanned for the same kinds of reusable material:

  • Ontology: what kinds of state or entity the architecture says must be distinguished.
  • Dynamics: how those entities become active, compete, cooperate, learn, or cause action.
  • Method: how the architecture turns a verbal idea into a model that can be inspected or tested.
  • Transfer candidates: concrete changes the idea might motivate in Commonplace or in agent-memory evaluation.
  • Non-transfer and failure modes: where the analogy is likely to break or create needless machinery.
  • Grounding questions: the smallest primary-source checks needed before attribution or adoption.

Candidate identifiers such as ACTR-1 are workshop-local handles. They do not assert stable vocabulary or graph relations.

What would close this workshop

The workshop closes when:

  1. the promising cross-architecture candidates have been deduplicated against the existing KB;
  2. candidates that could materially alter a design or evaluation have either been grounded in primary sources or rejected without that expense;
  3. durable conclusions have been extracted into the appropriate notes, references, instructions, or proposals; and
  4. the remaining architecture summaries no longer carry live work.

At closure this directory should be deleted, as normal for the workshop layer.

Existing Commonplace anchors

Complete file listing (generated at build time)