KurisutinaEinstein, G. O., McDaniel, M. A., Manzi, M., Cochran, B., & Baker, M. Prospective memory and aging: Forgetting intentions over short delays. Psychology and Aging, 15(4), 671–683. Peer-reviewed primary human experiments. Canonical DOI; PubMed; author-uploaded full text.
Factual core
Three experiments tested responding to an encountered target word immediately or during a later task phase. Experiment 1 enrolled 40 younger and 40 older adults; salient uppercase targets supported near-ceiling immediate responses, while delayed execution declined and divided attention reduced delayed performance. Delays were 39 seconds for younger adults and 30 for older adults because ongoing-task pacing differed. Experiment 2 enrolled 24 adults per age group, using 10- or 30-second filled or unfilled delays: older adults responded less often; longer delays did not significantly worsen performance, and unfilled delays did not benefit older adults. There was no immediate-response arm. Experiment 3 enrolled 40 older adults: lowercase targets yielded low performance without a significant additional delay decrement. Initial retrieval was inferred from controls and cue salience, not separately observed on delayed trials. Working-memory associations and several explanatory accounts were reported; the experiments did not identify a unique mechanism. Full article
Inferential audit — our analysis
The measured endpoint combines noticing the cue, accessing the instruction, retaining or recovering the pending intention, recognizing the permitted opportunity, and executing the action. The successful immediate controls make a downstream problem plausible, but do not identify the failing stage on any particular delayed trial. A separate detection response would still not establish retrieval of the correct action; an action-content check would be stronger but could itself reinforce maintenance. This distinction matters when designing an extraction test.
The most tempting overclaim is that difficult retrieval protects a successfully retrieved intention. Experiment 3 compares final response rates between different groups; it does not measure the fraction of confirmed retrievals that survive. Its low performance and small arms leave room for floor compression, changes in monitoring or strategy, and different mixtures of successful retrieval. A nonsignificant delay effect establishes neither equivalence nor universal preservation. Salience also differs between experiments rather than being crossed with delay in the same experiment.
Likewise, an apparently unfilled interval is not a measurement of participants' thoughts. The findings cannot isolate passive decay from distraction, spontaneous recovery, prospective replanning, or opportunity recognition. Different pacing and response support across age groups help accommodate task difficulty but complicate a pure aging interpretation; matching task performance would still not prove matching effort. Span correlations and regression attenuation are not interventions on working memory and do not establish causal mediation. The various accounts should remain competing hypotheses.
Proposed functional recipient test — original proposal
Extract a participant's newly formed intention before a held-out opportunity and supply it to a recipient system. Separately vary cue salience, interruption, delay, opportunity boundaries, and later cancellation or completion. Score whether the recipient performs the appropriate action at the permitted opportunity, omits it, acts prematurely, or repeats a completed action. Ask for the remembered instruction only after the action window, with separate probe and no-probe arms where measurement reactivity matters.
Compare equivalent source information represented as a narrative instruction, explicit trigger–action–opportunity fields, and those same fields with persistent pending/completed/cancelled state. Hold semantic content and external reminders constant across representation comparisons; an oracle pending-state flag would add information rather than merely change its format. Introduce a fresh intention after transfer to test whether the recipient can update the representation, rather than only replay an imported instruction. Include a simple task scheduler as a baseline for successful execution. If person-specific forecasting is also desired, score that separately against the participant's observed behavior; reliable execution and reproducing a person's omission pattern are different targets.
The implication is an architectural hypothesis: extracting an intention's verbal content may be insufficient unless the recipient also supports triggering, deferred execution, and state updates. This study tests human task performance, not transfer between people and systems, and does not determine which machine representation will work.
Access and reading scope
Accessed 2026-09-22. Read the complete main introduction, all three methods/results sections, general discussion, and extracted table/caption text. The source covers all article text from its title through acceptance dates, with no missing source lines in that interval. References were retained but not independently investigated. No supplement was identified. Figures and table layout were not visually audited; no graph values were digitized.
The public author-uploaded ResearchGate extraction has poor OCR, especially split final letters and flattened tables. Direct ResearchGate HTML/PDF and the publisher PDF returned HTTP 403; no PDF was verified or retained. The archival source is therefore the web reader's line-indexed extraction, not original HTML or a downloaded PDF. The readable derivative only collapses whitespace; it does not silently repair words or numbers. Numerical conclusions above rely on legible prose, not reconstructed graph heights or ambiguous table cells. Public accessibility does not establish an open redistribution license.
Crossref identifies the single-slash DOI as canonical and lists 10.1037//0882-7974.15.4.671 as an alias; PubMed uses that alias. Local artifacts: source extraction, readable text, Crossref record, provenance and hashes.