Memory Distortion in Cognitive and AI Systems

• Memory distortion is a phenomenon characterized by systematic deviations between actual events and remembered details due to constraints like finite memory and interference.
• It spans multiple fields—from cognitive experiments on false recollection to algorithmic studies in AI and information theory—highlighting varied impacts on decision quality and representation.
• Applications include optimizing caching systems, improving agent memory design, and understanding biases in human memory, offering actionable insights into compression and fidelity trade-offs.

Memory distortion denotes a family of phenomena in which remembered, represented, or reconstructed information departs systematically from an objective baseline because of finite memory, post-event inputs, causal encoding constraints, or memory-dependent fidelity criteria. In recent literature, the term spans false recollection induced by AI-edited media, similarity-induced bias in visual working memory, decision-quality loss caused by compressing histories into a bounded set of runtime states, and information-theoretic trade-offs in which memory resources alter achievable distortion, subjective time, or reward valuation (Pataranutaporn et al., 2024, Cao et al., 14 Jul 2025, Zou et al., 11 May 2026, Ortega et al., 2016, Abin et al., 29 Jan 2026). This suggests that the unifying issue is not simply inaccuracy, but which distinctions a system preserves, merges, or reweights under constraint.

1. Conceptual scope across cognitive science, AI, and information theory

The modern literature uses “memory distortion” in at least four technically distinct senses. In human-memory experiments, it refers to systematic deviation between what is recalled and what actually occurred or was originally experienced. In agent-memory research, it denotes the loss in achievable decision quality induced by compressing many histories into a few reusable runtime states. In communications and caching, it refers to fidelity loss—typically squared-error distortion—under explicit memory budgets. In rate–distortion theory with memory, it denotes settings in which distortion depends on past reproductions, so the fidelity criterion itself has memory.

Domain	Memory object	Distortion object
Human recollection	episodic record of an original image or event	false recollection, misattribution, suggestibility
Subjective time and choice	predictive dependencies between past and future sensorimotor interactions	renormalized subjective time axis and reward discounting
Agent memory	$K$ reusable runtime states	loss in achievable decision quality
Communications and caching	cache memory, causal state, or past reproductions	MSE or memory-dependent distortion

Despite their heterogeneity, these formulations share a structural theme: memory is not a passive archive but a constrained representational resource. In one class of models, the constraint is biological or cognitive; in another, it is algorithmic; in another, it is a formal budget in bits, states, or channel rate. A plausible implication is that “memory distortion” is best understood as a systems-level phenomenon in which compression, interference, and causal structure determine what information remains operationally available.

2. Coding-efficient memory, subjective time, and intertemporal valuation

A particularly influential formalization treats memory distortion as a consequence of coding efficiency in sensorimotor representation. In this framework, agent–environment interaction is a stochastic process $P$ generating action–observation sequences, and the agent maintains a minimal sufficient statistic $s$, the “present,” containing precisely the information in the past that helps predict the future (Ortega et al., 2016). The central quantity is predictive information,

$$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$

which measures the minimal bits of memory linking past to future.

Within this account, subjective duration is not identified with clock time but with the amount of re-encoding required when new interaction data change the present. The pointwise “present” is

$$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$

and the perceived duration of a new interaction $x_{\mathrm{n}}$ is

$$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$

Events that strongly constrain future behavior therefore require more bits to encode and are perceived as longer, whereas predictable or behaviorally irrelevant interactions produce little change in $s$ and feel short.

The same framework defines subjective time over a window of physical time $t$ as cumulative predictive information,

$$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$

Different probabilistic model classes induce different growth laws. For a finite model, $P$0; for a parametric model, $P$1; for a nonparametric model, $P$2 with $P$3. All are sublinear in physical time, so the subjective time axis is renormalized: early, information-rich periods dilate, whereas later periods compress as predictive dependencies saturate.

Valuation is coupled to the same information constraint through a free-energy control objective,

$P$4

Its optimizer is a Gibbs posterior over futures, and the resulting change in choice probability is proportional to “rejoice” or negative regret. Because predictive information grows differently across model classes, discounting laws differ as well: finite models induce linear decay to zero, parametric models induce exponential discounting in subjective time, and nonparametric models induce hyperbolic discounting. Mapping subjective time back to physical time can transform exponential discounting in $P$5 into hyperbolic discounting in $P$6. The framework therefore unifies temporal distortion and intertemporal preference by a single coding-efficiency principle: interactions that most constrain future behavior feel longer and are assigned greater current value.

3. Decision-centric memory distortion in bounded agents

A more recent formulation relocates memory distortion from perceptual timing to bounded decision-making. Histories $P$7, queries $P$8, actions $P$9, and reward means $s$0 define a setting in which the full-history oracle value is $s$1, while the answer-time memory state is a discrete variable $s$2 produced by a query-aware encoder $s$3 (Zou et al., 11 May 2026). Distortion is explicitly decision-level: $s$4 Memory is valuable not because it reconstructs the past, but because it preserves the distinctions between histories that must remain separate under a fixed budget to support good decisions.

This perspective yields an exact forgetting boundary. For a fixed query $s$5, tolerance $s$6, and nonempty set $s$7, safe merging is equivalent to the existence of a single action $s$8 with $s$9 for all $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$0. Equivalently, a one-state encoder and deterministic decision rule can achieve worst-case distortion at most $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$1 on $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$2 if and only if the histories in $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$3 share a common $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$4-optimal action. Descriptive similarity is therefore neither necessary nor sufficient for safe forgetting. This is sharpened by the cluster decision radius

$$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$5

and the pairwise decision distance

$$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$6

The budgeted trade-off is summarized by the worst-case frontier

$$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$7

with corresponding covering and packing numbers $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$8 and $$I(X_{\mathrm{p}};X_{\mathrm{f}})=\mathbb{E}_{P}\!\left[\log\frac{P(X_{\mathrm{f}}\mid X_{\mathrm{p}})}{P(X_{\mathrm{f}})}\right],$$9. Any stochastic memory supporting worst-case distortion at most $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$0 must satisfy

$$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$1

where $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$2 is uniform over a maximum packing. The paper also states that deciding whether $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$3 is NP-complete, which rules out exact global optimization in the deterministic $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$4-slot setting.

The resulting algorithm, DeMem, implements conservative online refinement. It maintains counts $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$5, empirical means $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$6, confidence radii

$$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$7

one-sided lower bounds on pairwise decision distance, and cluster-radius certificates. It builds a cannot-link graph over observed contexts, performs binary search over conflict thresholds, and uses greedy degeneracy coloring to produce a feasible $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$8-way partition; per-epoch time is $$\text{Present}(x_{\mathrm{f}})=\log\frac{P(x_{\mathrm{f}}\mid x_{\mathrm{p}})}{P(x_{\mathrm{f}})}=\log\frac{P(x_{\mathrm{f}}\mid s)}{P(x_{\mathrm{f}})},$$9. With probability at least $x_{\mathrm{n}}$0,

$x_{\mathrm{n}}$1

while the minimax lower bound is

$x_{\mathrm{n}}$2

Empirically, DeMem attained the best overall LoCoMo accuracy on both backbones, with $x_{\mathrm{n}}$3 for GPT-4o-mini and $x_{\mathrm{n}}$4 for GPT-4.1-mini; descriptive retrieval recalled approximately $x_{\mathrm{n}}$5–$x_{\mathrm{n}}$6 of gold evidence under matched budget, whereas DeMem reached $x_{\mathrm{n}}$7, compared to an oracle at $x_{\mathrm{n}}$8. In this literature, memory distortion is the irreducible decision loss created by compressing histories more aggressively than the action structure permits.

4. Causal rate–distortion and distortion functions with memory

In information theory, memory distortion arises in two related but distinct ways. One concerns sources and channels that have memory, so realizable coding must be causal. The other concerns fidelity criteria that themselves depend on past reconstructions. Both settings depart from the classical memoryless rate–distortion paradigm (Kourtellaris et al., 2013, Kourtellaris et al., 2013, Abin et al., 29 Jan 2026).

For sources with memory, the key object is the nonanticipative reproduction distribution

$x_{\mathrm{n}}$9

which enforces that reproduction at time $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$0 depends only on current and past source symbols and past reproductions. The associated nonanticipative rate–distortion function is

$$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$1

where directed information replaces the classical mutual-information objective. The causal formulation is equivalent to Gorbunov–Pinsker nonanticipatory $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$2-entropy, and under stationarity admits an exponential-tilt Gibbs form for the optimal kernel. Achievability of symbol-by-symbol coding with memory without anticipation follows when the optimal causal reproduction is realizable by an encoder–channel–decoder cascade and the matching condition $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$3 holds.

The case study of a Binary Symmetric Markov Source $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$4 under Hamming distortion over a Binary State Symmetric Channel yields exact matching. For $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$5,

$$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$6

with $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$7, $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$8, and $$\text{Duration}(x_{\mathrm{n}})=\log\frac{P(x_{\mathrm{n}}\mid x_{\mathrm{f}})}{P(x_{\mathrm{n}}\mid x_{\mathrm{p}})}.$$9. Setting the channel parameters $s$0, $s$1, and cost $s$2 makes channel capacity equal to $s$3. Under that alignment, the identity mapping $s$4 and $s$5 is optimal, and the matched symbol-by-symbol scheme achieves average distortion $s$6. The papers also provide Hoeffding-type bounds for Markov chains showing exponential decay of excess distortion probability.

A different generalization places memory not in the source but in the distortion measure. “Subjective distortion” defines the per-block average distortion as

$s$7

with fixed initial symbol $s$8. The source is i.i.d., but the instantaneous fidelity cost depends on the current source symbol, the current representation, and the previous representation. This yields a multi-letter noncausal characterization via Han’s information-spectrum theorem and single-letter bounds. The Markov-kernel inner bound uses a stationary kernel $s$9 and achieves

$t$0

while the simpler memoryless-kernel bound is

$t$1

Outer bounds are obtained through an independent-copy replacement and a relaxation to $t$2 subject to the same average distortion. When $t$3 reduces to a memoryless distortion $t$4, the upper and lower bounds coincide with the classical rate–distortion function. These results show that distortion may inherit memory either from the source-channel architecture or from the fidelity criterion itself.

5. Cache memory, effective rate, and reconstruction distortion

In caching and wireless video delivery, memory distortion has an explicitly engineering meaning: finite cache memory and finite delivery rate together determine the reconstruction distortion at the receiver. The common Gaussian source model treats each file $t$5 as successively refinable under squared-error distortion, with file-specific distortion–rate function

$t$6

so every additional bit per sample in effective rate reduces MSE exponentially (Hassanzadeh et al., 2019, Hassanzadeh et al., 2015).

The central quantity is effective rate. In Local Cache-aided Unicast (LC-U), the effective rate seen by user $t$7 for file $t$8 is $t$9, where $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$0 is cache memory devoted to the file and $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$1 is delivered unicast refinement. Placement solves the convex reverse-water-filling problem

$$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$2

with solution

$$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$3

Delivery for a realized demand $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$4 is another reverse-water-filling problem across users,

$$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$5

Cooperative Cache-aided Coded Multicast replaces pure unicast refinement with a split $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$6, where $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$7 is the coded multicast portion and $$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$8 is residual uncoded enhancement. The storing range

$$\tau_s(t)\equiv I\bigl(X_{\mathrm{p}}^{(t)};X_{\mathrm{f}}^{(t)}\bigr).$$9

specifies the cumulative rate guaranteed by coded multicast for file $P$00 to user $P$01, and the packet caching probability is

$P$02

The paper specializes coded delivery to RF-GCC, obtains closed-form upper bounds $P$03 and $P$04, and uses these bounds to optimize memory allocation and storing ranges under heterogeneous cache sizes, popularity distributions, and file variances.

This framework formalizes a distortion–rate–memory region. A tuple $P$05 is achievable if $P$06 and $P$07, with optimal distortion–rate–memory function

$P$08

The key mechanism is additive in rate but exponential in distortion: cache memory contributes linearly to effective rate, whereas distortion decays as $P$09. Coded multicast improves the aggregate effective rate delivered under the same broadcast budget and thereby lowers expected distortion significantly beyond unicast baselines.

The numerical consequences are substantial. In the heterogeneous case with Zipf $P$10, $P$11, $P$12, and $P$13, CC-CM reduces expected distortion by approximately $P$14 versus LC-U at $P$15 and $P$16, and by approximately $P$17 at $P$18 and $P$19. In the uniform case with $P$20 and $P$21, the reduction is approximately $P$22 at $P$23 and $P$24, reaching approximately $P$25 at $P$26. In this literature, “memory distortion” is not mnemonic error but the fidelity penalty or improvement produced by explicit memory-rate allocation.

6. False recollection from AI-edited images and videos

In experimental human-memory research, memory distortion retains its classical sense: any systematic deviation between what is recalled and what actually occurred or was originally experienced. A recent pre-registered study operationalized a false memory as an incorrect answer to a masked-detail question about an original image after exposure to AI-altered material (Pataranutaporn et al., 2024). The paradigm strongly implicates source misattribution, because participants first saw original images and later saw edited images or generated videos before being asked about the originals.

The study recruited $P$27 U.S.-based adults, randomized $P$28 per condition, with gender balanced $P$29 female:male and age $P$30–$P$31 ($P$32, $P$33). Participants viewed $P$34 original images for two minutes, completed a two-minute Pac-Man filler, and were then exposed to one of four conditions: unedited images, AI-edited images, AI-generated videos of unedited images, or AI-generated videos of AI-edited images. Adobe Photoshop AI was used to modify $P$35 of the $P$36 images, and Luma’s Dream Machine generated $P$37-second videos. All altered stimuli were labeled “AI-enhanced image.”

The primary outcome was false memories over the $P$38 targeted items. Mean false-memory rates were $P$39 in the control condition, $P$40 for AI-generated videos of unedited images, $P$41 for AI-edited images, and $P$42 for AI-generated videos of AI-edited images. Shapiro–Wilk indicated non-normality, and a one-way Kruskal–Wallis test yielded $P$43, $P$44. Dunn post hoc tests with FDR correction showed significant differences for control versus AI-edited images ($P$45), control versus AI-generated videos of AI-edited images ($P$46), and AI-generated videos of unedited images versus AI-generated videos of AI-edited images ($P$47). Relative to control, AI-edited images increased false recollections by $P$48, AI-generated videos of unedited images by $P$49, and AI-generated videos of AI-edited images by $P$50.

Correct memory declined correspondingly. Mean non-false memory rates were $P$51 in control, $P$52 for AI-generated videos of unedited images, $P$53 for AI-edited images, and $P$54 for AI-generated videos of AI-edited images. Confidence in false memories also increased: mean confidence on a $P$55–$P$56 scale was $P$57 in control, $P$58 for AI-edited images, and $P$59 for AI-generated videos of AI-edited images, with Kruskal–Wallis $P$60, $P$61. The weighted score combining accuracy and certainty dropped sharply from $P$62 in control to $P$63 in the video-of-edited condition; one-way ANOVA gave $P$64, $P$65, with partial $P$66. Among measured moderators, age had a small negative coefficient, $P$67, $P$68, $P$69, $P$70, implying slightly lower odds of false memories with increasing age.

The proposed mechanisms are familiar from the misinformation literature but updated for generative media. The paper identifies perceptual realism and fluency, source monitoring errors, gist strengthening under Fuzzy-trace theory, imagination inflation, and the misinformation effect as plausible drivers. Passive labels were not sufficient to prevent distortion: despite “AI-enhanced” disclosure, false memories and confidence rose. The strongest distortions were observed for “AI-generated videos of AI-edited images,” especially for people and environmental edits. The study therefore places memory distortion at the intersection of synthetic media, HCI, and eyewitness reliability, with direct implications for provenance tracking, interface warnings, and evidentiary standards.

7. Similarity-induced distortion in visual working memory

A separate experimental tradition studies distortion of visual working memory during retention. Here, memory distortion is a retroactive change in the stored representation of a target item induced by intervening perceptual input and tasks, operationalized as a shift in the reported item on a circular stimulus wheel away from the original target and toward the compared item (Cao et al., 14 Jul 2025). The paper uses the term similarity-induced memory bias (SIMB) for this attraction-like effect and distinguishes it from slow drift, generic interference, and categorical bias.

The experiments introduce an AI-driven framework for generating naturalistic object stimuli from behavior-based dimensions. Forty-two dimensions were categorized as visual, semantic, and mixed. “Image wheels” were formed by smoothly editing the activation of two selected object dimensions in the latent space of a generative model while preserving base-image identity and appearance; “dimension wheels” were generated directly from dimension activation values without requiring visual similarity to the base image. The underlying controllable generation framework adopts a two-stage strategy, first modeling $P$71 and then $P$72, where $P$73 denotes conditioning parameters, $P$74 CLIP embeddings, and $P$75 the generated image. The wheel construction uses dimensional guidance,

$P$76

smoothness guidance,

$P$77

and, for image wheels, CLIP guidance,

$P$78

The behavioral paradigm used set size $P$79. The memory item was shown for $P$80 ms, followed by a $P$81 ms retention interval, two same–different similarity judgments during retention, a $P$82 ms delay, and then wheel-based report. The Image Wheels Induction Experiment involved $P$83 participants and $P$84 trials per participant, comprising $P$85 baseline trials and $P$86 induction trials. The Dimension Wheels Induction Experiment involved $P$87 participants with matched procedure and trial counts. Low-confidence trials were excluded, baseline accuracy below $P$88 led to participant exclusion, and induction trials were counted as valid only if the participant selected the induction item in both similarity judgments.

The principal metric was the bias score, printed as

$P$89

that is, the unsigned angular deviation between reported and target wheel positions. The main result is that perceptual comparison increased memory distortion relative to baseline in both experiments. Across wheel types, image wheels produced larger distortions than dimension wheels, with mean accuracy $P$90 and mean bias $P$91 for image wheels, versus accuracy $P$92 and mean bias $P$93 for dimension wheels. This indicates that holistic visual similarity is a strong driver of distortion, but dimension-level similarity alone is sufficient.

Dimension category further modulated vulnerability. In the Image Wheels Induction Experiment, visual dimensions yielded accuracy $P$94 and bias $P$95, whereas semantic dimensions yielded accuracy $P$96 and bias $P$97. In the Dimension Wheels Induction Experiment, visual dimensions yielded accuracy $P$98 and bias $P$99, whereas semantic dimensions yielded accuracy $s$00 and bias $s$01. Mixed dimensions occupied intermediate positions, and mixed+semantic distortion dropped to $s$02 in the dimension-wheel condition. The paper reports no inferential statistics, confidence intervals, or mixture-model fits, but the descriptive pattern is consistent: visual dimensions are more prone to comparison-induced distortion than semantic dimensions.

The proposed mechanism is representational overlap. Comparing a stored target to a similar input engages overlapping feature codes and increases the weight of shared features, thereby biasing the stored representation toward the inducer. Semantic organization may stabilize memory through schema-based priors and deeper processing, reducing vulnerability. In this account, memory distortion is neither random error nor mere overload; it is a structured displacement along behaviorally meaningful similarity dimensions.