ASMO: transferable scan

Status: memory-first and source-ungrounded Recall confidence: low

Remembered model

I remember ASMO chiefly as an attention-centered cognitive architecture used in robotic or situated systems. I do not trust my recall of the acronym expansion, module inventory, or learning rules. The stable-looking theme is selection among competing perceptual, cognitive, and goal-relevant demands under bounded processing.

This scan therefore extracts only from that broad theme. It intentionally does not attribute a specific salience equation or component structure to ASMO.

Provisional ontology

The concepts worth checking appear to be:

  • Attention candidate: something that could receive scarce processing now.
  • Priority signal: a contribution from novelty, urgency, current goals, learned relevance, or persistent commitments.
  • Selector: the arbitration process that combines signals and commits resources.
  • Attended item/process: the winner or small active set.
  • Inhibition or persistence: mechanisms preventing either thrashing or permanent capture.
  • Attention shift: a transition that should have a reason visible in the execution trace.

The important distinction is between an item's content and the reason it wins attention. A highly authoritative rule, a surprising observation, and a user-designated goal can all deserve priority for different reasons.

Transfer candidates

  • ASMO-1 — make context selection inspectable. A context engine should be able to report which candidates competed, which priority sources mattered, and why a candidate was excluded. Similarity rank alone is not an attention explanation.
  • ASMO-2 — keep priority channels typed. Novelty, goal relevance, risk, authority, and recency should not become one uninterpretable score unless their trade-offs are explicit. Typed channels permit policy changes and postmortems.
  • ASMO-3 — add persistence and interruption rules. Agent loops need a principled balance between staying with the current inquiry and reacting to newly salient evidence. This could become a control-policy vocabulary: interrupt, queue, decay, or ignore.
  • ASMO-4 — evaluate attention by opportunity cost. A selector is not good merely because chosen items are relevant; evaluation should include important candidates that lost and the work displaced by the winners.

These are candidate refinements to designing agent memory from consumer effects: memory activation needs an arbitration story when more useful material is available than can be loaded.

Method worth borrowing

An attention architecture encourages perturbation tests. Hold the available information constant while varying the goal, novelty, urgency, or persistence signal; then predict which item becomes active and what action changes. That is more diagnostic than measuring retrieval precision against a context-free relevance label.

Non-transfer and failure modes

  • The remembered attention theme may omit ASMO's actual distinguishing mechanism.
  • A centralized selector can become a fictitious homunculus unless its inputs and decision rule are implemented.
  • A single priority score can conceal normative choices about whose goal or risk dominates.
  • Robot sensor salience may transfer poorly to text-heavy knowledge work.

Grounding questions

  1. What does ASMO stand for, and what problem was it explicitly designed to solve?
  2. Is attention centralized, distributed, learned, or rule-based?
  3. Which priority sources does the architecture distinguish?
  4. What robotic or non-laboratory results actually depend on the attention mechanism?