Memory research visualization

Interactive chart

Two interactive views. Lifespan shows trajectories of each memory system across age 20–85 with demographic subgroup comparison. Encoding & retention shows how much of a newly-encoded memory remains over time, under different practice and condition tiers. All curves are stylized illustrations of patterns from the literature, not empirical data.

Project goal

Render the differential aging trajectories of distinct human memory systems across the adult lifespan, and let the user toggle between demographic groupings to see how each subgroup modifies the curves. The artifact is built to be a starting point for iteration: replace stylized curves with empirical data, expand the systems shown, expand the demographic dimensions, and improve the presentation.

Research synthesis

Memory taxonomies relied on

• Atkinson & Shiffrin (1968) multi-store model — sensory / short-term / long-term tiers.
• Baddeley & Hitch (1974, updated 2000) working memory — central executive, phonological loop, visuospatial sketchpad, episodic buffer.
• Tulving / Squire long-term memory taxonomy — declarative (episodic, semantic) vs non-declarative (procedural, priming, conditioning, non-associative).
• Cowan's embedded-processes model — working memory as activated LTM plus focus of attention (~4 items).
• Craik & Lockhart (1972) levels-of-processing — durability depends on encoding depth, a process counterpoint to the structural models.

Differential aging pattern (general population baseline)

Theoretical anchors: Salthouse processing-speed theory; Park & Reuter-Lorenz scaffolding theory of aging and cognition (STAC / STAC-r); Cattell-Horn fluid vs crystallized distinction maps onto fluid-like systems (WM, processing speed, episodic) vs crystallized-like (semantic).

Demographic dimensions in the widget

1. General population — baseline curves above.
2. Education (low ≤HS / college / graduate) — cognitive reserve (Stern). Higher education raises baseline and flattens decline, especially for WM and episodic; small effect on procedural and semantic. Confounded with SES.
3. Sex (female / male average) — small effects, d ≈ 0.2–0.5, heavy overlap. Female advantage on verbal episodic recall; male advantage on visuospatial WM. Verbal-memory advantage in females may delay AD detection on standard tests.
4. Childhood SES (low / medium / high) — hippocampal and prefrontal sensitivity to early environment (Noble, Farah, Hackman). Persists into adulthood. Strongest on WM and episodic.
5. Adult SES (low / medium / high) — ongoing effects via occupational complexity (Schooler), cognitive engagement, healthcare access, and stress exposure. Effect concentrated in late-life decline rate rather than early-life baseline. Confounded with education and childhood SES; modeled here as net of those.
6. Bilingualism (monolingual / bilingual) — modest reserve effect on executive control and WM, mainly in older adulthood. Actively contested (Bialystok positive findings vs Paap critiques) — widget shows the optimistic end of the range.
7. Physical exercise (sedentary / active) — aerobic fitness associated with preserved hippocampal volume and slower WM/episodic decline (Erickson, Kramer). Reliable in midlife and later.
8. Diabetes (none / type 1 / type 2) — modest cognitive deficits in T1D (Brands et al. 2005; Brinkman et al. 2012 meta-analyses), strongest in processing speed and verbal memory; severe-hypoglycemia history and childhood onset are the main moderators. T2D shows larger effects, accelerated late-life decline, and elevated dementia risk via vascular and metabolic pathways (Biessels; Whitmer).
9. Cognitive engagement (low / moderate / high) — sustained reading, puzzles, and novel learning build reserve (Stern); high engagement flattens decline across episodic and semantic memory, magnitude comparable to formal education (Wilson 2007, 2013).
10. Sleep quality (chronic short / normal / good) — slow-wave sleep supports episodic consolidation (Walker); chronic short sleep accelerates episodic and processing-speed decline, effect concentrated in late life (Lim & Yaffe 2013; Spira 2013).
11. Media multitasking (light / heavy) — Ophir, Nass & Wagner (2009) reported attention-filtering deficits in heavy multitaskers, but later replications and a 118-study meta-analysis find only a small, age-uncertain effect on working memory and attention (Wiradhany 2017) — magnitudes here are intentionally tiny.
12. Hearing loss (none / mild-moderate / severe) — Lancet Commission 2020 ranks midlife hearing loss the largest modifiable dementia risk factor (Livingston); each 10 dB raises decline risk ~16–27%, hitting verbal-episodic and processing speed hardest (Lin 2011).
13. APOE ε4 status (non-carrier / heterozygous / homozygous) — clear allele-dose effect on age of decline onset, ~50s in homozygotes, ~60s in heterozygotes, ~70s in non-carriers (Caselli 2009); episodic memory is the predominant target (Liu 2013).
14. Clinical status (healthy / MCI / early Alzheimer's) — disease starts in medial temporal lobe, so episodic memory collapses first and most severely. Procedural relatively preserved even at moderate AD. MCI is increasingly treated as a clinical prodrome.

Longitudinal data sources that inform the curve shapes

• Seattle Longitudinal Study (Schaie)
• Baltimore Longitudinal Study of Aging
• English Longitudinal Study of Ageing (ELSA)
• Rönnlund et al. meta-analyses (Betula project)

Critical caveats

The curves are calibrated to reflect:

• the right shape (peak age, decline slope, whether the curve is monotonic),
• the right relative ordering of subgroups,
• the right relative magnitude of differential vulnerability.

Other caveats baked into the framing:

• Population averages; within-group variance is large and overlaps heavily between groups.
• Many demographic effects are confounded (education × SES × health behaviors).
• Standardized memory tests have known validity issues across languages and cultures.
• Cross-sectional designs (most older data) confound aging with cohort effects; longitudinal data is more trustworthy but slower to accumulate.

Implementation

Built as an HTML widget using Chart.js 4.4.1 (loaded from cdnjs). Two control groups (demographic dimension, memory type), a custom HTML legend, and a single line chart canvas. The memory-type selector is disabled when "general population" is active (general view shows all five memory types overlaid).

Curves are computed in two ways:

• Linear transforms for subgroups whose effect is well-modeled as a baseline offset plus a decline-slope multiplier (education, SES, bilingualism, exercise). transform(curve, offset, slope) applies offset everywhere and steepens or flattens the post-peak decline.
• Custom curves for subgroups whose effect is qualitative or onset-timed (sex, clinical status). Clinical curves use a decline(startIdx, rate) helper that keeps the curve identical to baseline until a start index, then subtracts a linear ramp.

Iteration directions

Priority order is roughly: replace stylized data → expand systems → expand demographics → improve presentation.

Replace stylized data with empirical points

• Overlay actual data points from Park et al. (2002) "Models of visuospatial and verbal memory across the adult life span" — clean cross-sectional curves for WM, episodic, processing speed across age that can be normalized to peak.
• Use Rönnlund et al. (2005) Betula longitudinal data for episodic and semantic — rare advantage of being longitudinal so it disentangles cohort effects.
• For clinical curves, use ADNI public data for healthy / MCI / AD trajectories on standardized neuropsych batteries.
• Add confidence-interval shading (typically ±1 SD or 95% CI bands) to communicate within-group variance, since the current single lines hide overlap.

Expand the memory systems

• Add prospective memory — lab vs naturalistic dissociation is interesting (Henry et al. meta-analyses).
• Add autobiographical memory specificity — declines with age, more so in depression and PTSD (Williams's overgeneral memory).
• Add source memory as a sub-component of episodic — particularly age-vulnerable.
• Consider splitting working memory into central executive vs phonological loop vs visuospatial sketchpad — they age differently.

Expand the demographic / event dimensions

• Early life adversity / ACEs — separate from SES, has distinct effects on hippocampus and HPA axis.
• Traumatic brain injury — concussion history, severity tiers.
• Sleep quality — strong effect on consolidation; could show as chronic-short-sleep vs adequate.
• Hearing loss — emerging Lancet Commission risk factor for dementia.
• Chronic stress / cortisol exposure — hippocampal effects.
• Hormonal transitions — menopause and verbal memory; testosterone and spatial.
• APOE ε4 carrier status — biggest single genetic risk for late-onset AD; well-suited to a 3-tier (non-carrier / heterozygous / homozygous) view.
• Cross-cultural — Wang's work on autobiographical memory differences; literacy effects on serial-position curves.

Earlier age range

Currently starts at 20. Adding 5–20 would let you show developmental trajectories — working memory and episodic memory both develop into the late teens, and the developmental SES effects are largest in childhood. Procedural and semantic curves would also become more informative.

Presentation improvements

• Add a "highlight one series, dim the rest" mode on hover.
• Add a comparison mode: pick two demographics and see the chart side-by-side or overlaid.
• Add a "what's happening neurally" tooltip for each memory type that explains the underlying brain systems (hippocampus for episodic, prefrontal for WM, basal ganglia and cerebellum for procedural, etc.).
• Surface the citations as clickable links per claim, so the user can audit which curve is grounded in which study.
• Make the dynamic caption / interpretation text part of the widget (current version puts interpretation in surrounding prose, which doesn't travel with the artifact).

Known limitations to address

• The transform() function is symmetric around peak — but real demographic effects often aren't (e.g., education mainly affects late-life decline, not young-adult baseline). Consider replacing the linear offset+slope model with a piecewise function or a parametric curve (sigmoid decline with shift, slope, and floor parameters).
• The clinical onset ages are hardcoded to mid-50s/60s; real MCI/AD onset distributions are wide. Consider letting the user shift onset age.
• Sex differences shown are global averages — real differences vary by task and by age (e.g., post-menopausal verbal memory changes).