The relationship between language learning technology and human cognition has long been framed around the limitations of working memory—the cognitive system responsible for the temporary storage and manipulation of information that underpins complex comprehension and production (Baddeley, 2000). In second language acquisition, working memory is widely recognized as a core bottleneck: tasks that overtax working memory tend to reduce accuracy, depth of processing, and long-term retention, particularly for learners with lower working memory capacity (Gathercole and Alloway, 2008). Consequently, the design of technology-enhanced language learning environments has historically been guided by the imperative to manage, or at least not exceed, this limited cognitive resource.

This review argues that the field is now in the midst of a transition between two technological eras with fundamentally different implications for working memory. The first is a pre-AI “Interactive Era,” in which technologies such as computer-assisted language learning software, hypermedia environments, captioned video, and online platforms provided valuable interactivity but were fundamentally static and pre-scripted (Lobin and Rösler, 2012). The second is an emerging “AI-Mediated Era,” in which generative AI chatbots, intelligent tutoring systems, biometric-adaptive platforms, and AI-orchestrated virtual and augmented reality environments are dynamic and dialogic, capable in principle of tailoring support to real-time learner states.

In the Interactive Era, interactivity was essentially reactive and bounded by pre-authored content, fixed branching, and rule-based feedback. Computer-assisted language learning software and online language courses grew in popularity, offering learners a range of interactive and engaging activities including multimedia presentations, language games, and virtual conversations (Jiang et al., 2017; Jones et al., 2017). However, these environments also presented challenges related to working memory: learners needed to manage cognitive load, cope with distractions, and navigate individual differences in working memory capacity to effectively learn new vocabulary, grammatical structures, and communication skills.

Within this paradigm, working memory was conceptualized primarily as a fixed individual-difference variable—a learner property that moderated the extent to which individuals could benefit from complex multimedia and self-directed tasks. Research in this era, especially under the influence of Cognitive Load Theory (Sweller et al., 2019), focused on mitigating the risk of overload by managing information density, controlling redundancy, sequencing tasks appropriately, and coordinating modalities to support dual-channel processing (Mutlu-Bayraktar et al., 2019). The underlying assumption was that good instructional design should stay within working memory limits; working memory itself was not something the system could sense, measure, or adapt to in real time.

Before proceeding, it is essential to distinguish among three theoretically and empirically distinct WM constructs that are often conflated in the literature (Unsworth and Engle, 2007; Gathercole and Alloway, 2008). WM capacity refers to the relatively stable, trait-like limit on the amount of information an individual can simultaneously maintain and manipulate, typically assessed through span tasks (e.g., digit span, reading span, operation span). WM utilization (or efficiency) refers to how effectively available WM resources are deployed under varying task conditions, including strategic allocation, attentional control, and resistance to interference—a process-level construct that can vary within individuals across contexts. WM training (or plasticity) refers to the potential for systematic practice to enhance WM capacity or efficiency, typically examined through extended training paradigms and transfer assessments. These three constructs carry distinct implications for instructional design: capacity sets an upper bound that instruction must respect, utilization determines how close to that bound learners can perform under given conditions, and plasticity determines whether instructional interventions can expand the bound itself. Critically, these distinctions have direct relevance for AI-mediated instruction, as AI systems might (a) respect capacity limits by dynamically adjusting task complexity, (b) improve utilization by optimizing scaffolding and reducing extraneous load, or (c) support plasticity through extended adaptive training regimens.

Empirical work in interactive environments repeatedly underscored this constraint. Varol and Erçetin (2021) showed that gloss type and position affected comprehension, with higher working memory capacity associated with better outcomes in hypermedia reading. Hong et al. (2021) reported that intrinsic cognitive load negatively affected flow experience, while gameplay interest positively affected flow in game-based environments. Türk and Erçetin (2014) found that interactive glosses were more effective than simultaneous glosses in promoting reading comprehension and incidental vocabulary learning, suggesting that better alignment with working memory limitations improves learning. Aryadoust (2020) reported that higher working memory capacity was associated with better performance on computerized while-listening tests. Across such studies, working memory consistently emerged as a predictor of success rather than a target for real-time intervention.

The arrival of powerful AI systems in language education is reshaping this landscape fundamentally. AI-mediated environments now include generative AI chatbots such as ChatGPT, Microsoft Copilot, and DeepSeek that can participate in open-ended dialogue, generate exemplars, and provide context-sensitive feedback. They also include adaptive learning platforms that adjust content difficulty, pacing, and support based on learner performance and, in advanced implementations, biometric indicators such as eye-tracking and physiological monitoring. Virtual and augmented reality environments combine immersive technologies with AI-driven content generation for multimodal, embodied language experiences (Squires, 2017). Intelligent personal assistants such as Amazon Alexa, Google Assistant, and Apple Siri enable voice-activated spoken interaction for listening and speaking practice.

These technologies are not limited to presenting pre-authored content; they can analyze learner input and behavior, detect patterns, and respond with adaptive scaffolding (Chen et al., 2020; Xu, 2025). This capability invites a reconceptualization of working memory from a static constraint to a dynamic target for adaptive compensation and regulation. Working memory becomes not only “something to design around” but potentially “something to design to”—a parameter that AI systems might measure, infer, and respond to. This evolution reintroduces the principle of Aptitude-Treatment Interaction, suggesting that AI systems could personalize instructional support based on a learner’s cognitive profile in ways previously impossible (Sana and Fenesi, 2025).

Initial empirical evidence points to both the promise and complexity of this technological shift. At a macro level, meta-analytic findings suggest that AI-assisted learning produces large overall effects (d = 1.17) relative to non-AI conditions, with particularly strong effects for vocabulary (d = 2.21) and receptive skills (d = 2.01). Notably, K–12 learners—whose working memory is still developing—show descriptively larger gains (d = 1.445) than college students (d = 0.988), a pattern consistent with a compensatory function of AI’s adaptive pacing for learners with less mature working memory capacity (Xu et al., 2025).

At a finer-grained level, recent AI-mediated interventions demonstrate that AI does not simply reduce cognitive load in the way traditional multimedia design sought to achieve; rather, it can redistribute load across working memory subsystems. Generative AI tools appear to offload lower-level encoding demands traditionally taxing the phonological loop while substantially increasing central executive demands for critical evaluation, prompt management, and integration of AI-generated content (Fan and Yao, 2025). Biometric-adaptive reading systems can preempt overload by adjusting difficulty in real time before errors accumulate (Yuan, 2025), while AI-driven multimodality enhances learning up to an empirically defined channel boundary of three to four concurrent sources, beyond which sensory overload undermines retention (Yu, 2025). AI-mediated data-driven learning can offload the search burden that traditionally taxed working memory in concordancing tasks, theoretically freeing resources for deeper noticing and internalization (Esfandiari and Allaf-Akbary, 2024).

Parallel evidence from neurocognitive research suggests that working memory itself may be more malleable than earlier static trait models implied. Computer-based multisensory working memory training using language materials has been shown to enhance both working memory capacity and second language ability, with transfer effects to novel auditory tasks and increased dorsolateral prefrontal cortex efficiency (Gkintoni et al., 2025). Additionally, evidence suggests that verbal, rather than spatial, working memory is particularly critical for second language achievement: high-achieving second language Chinese learners significantly outperformed low achievers on forward and backward digit span tasks but showed no differences in spatial working memory, with achievement correlating strongly with backward span (Xiao et al., 2025). These findings have direct implications for what aspects of working memory AI systems should prioritize scaffolding (Stakanova, 2023).

Despite these advances, a critical empirical gap tempers the promise of AI-mediated working memory support. Systematic reviews consistently document a measurement paradox: as AI in language education invokes working memory and cognitive load more explicitly, it measures working memory less directly. Chalmers et al. (2021), analyzing 111 s language aptitude studies spanning six decades, found working memory among the most frequently examined cognitive predictors (8.1% of independent variables) yet identified no AI-mediated interventions explicitly targeting or measuring working memory. Zhang and Aubrey (2024) found only one study examining working memory as an individual-difference factor across all second language pragmatics research, with no AI-mediated pragmatics studies including working memory measures. Goh and Aryadoust (2025) reported that while intelligent personal assistants and generative AI are increasingly used for second language listening and speaking, none of the reviewed AI studies included working memory tasks or aptitude-treatment interaction analyses. Xu (2025) noted that intelligent personal assistants are widely theorized to engage working memory through real-time communication practice, but direct empirical evidence on AI-mediated working memory outcomes remains absent.

Even the AI-mediated primary studies with the strongest cognitive focus rely on cognitive load scales and process indicators as proxies rather than direct working memory assessment. This creates a causal gap: without measuring working memory directly, researchers cannot determine whether AI reduces task demands (a design property), enhances functional working memory utilization (a processing property), or improves working memory capacity itself (a training property). Evidence that working memory may influence learning indirectly through proficiency (Teng, 2024) further complicates the picture, suggesting that working memory–AI relationships may involve mediation pathways that require sophisticated longitudinal and aptitude-treatment interaction designs to untangle.

Against this backdrop, the present systematic review is explicitly structured as a historical-comparative analysis of working memory in technology-enhanced language learning. To address the conceptual and empirical gaps identified above, the review is organized around three research questions:

Corresponding to these research questions, the review pursues three objectives. The first objective (aligned with RQ1) is to synthesize findings from a corpus of 27 empirical studies on the challenges and affordances of working memory in traditional, non-AI interactive language learning environments from 2010 to 2024, including computer-assisted language learning software, hypermedia, online platforms, and multimedia, thereby deriving evidence-based design guidelines for WM-sensitive instruction. The second objective (aligned with RQ2) is to juxtapose these findings with evidence from four recent AI-mediated primary empirical studies (Esfandiari and Allaf-Akbary, 2024; Fan and Yao, 2025; Yu, 2025; Yuan, 2025) that directly examine cognitive load and working memory-relevant processes in AI-assisted writing, adaptive reading, multimodal vocabulary learning, and AI-mediated data-driven learning, analyzing how AI affordances interact with WM subsystems in ways that differ from pre-AI technologies. The third objective (aligned with RQ3) is to identify the empirical boundary conditions under which AI-mediated support optimizes versus undermines WM functioning, and to document unintended consequences—particularly the cognitive load redistribution from encoding to evaluation, the measurement paradox whereby WM is increasingly invoked but decreasingly measured, and potential over-scaffolding (i.e., overprotection) with downstream implications for metacognitive monitoring and executive-skill development—that carry implications for future research and practice.

To systematize the relationship between WM constructs and AI affordances, Table 1 presents a conceptual framework mapping the three WM constructs (capacity, utilization, and training/plasticity) to four key AI affordances (adaptivity, multimodality, generative support, and feedback timing). For each pairing, the framework specifies the theorized mechanism and articulates a testable prediction suitable for future empirical investigation. For example, AI-driven adaptivity is hypothesized to respect capacity limits by dynamically adjusting task complexity, leading to the testable prediction that learners with lower WM capacity will show larger performance gains under adaptive versus fixed-difficulty conditions. Similarly, multimodal AI environments may improve WM utilization by distributing load across subsystems (phonological loop, visuospatial sketchpad), predicting that optimal retention will occur within a bounded channel range (e.g., 3–4 concurrent sources) with diminishing returns beyond this threshold. Generative AI support may offload lower-level encoding processes while increasing central-executive demands, predicting construct-specific load redistribution measurable via differentiated cognitive load instruments. Finally, preemptive feedback timing enabled by biometric AI may prevent utilization failures before they accumulate, predicting superior outcomes compared to reactive (post-error) adaptation. This framework enables researchers to formulate and test construct-specific hypotheses rather than treating WM as an undifferentiated variable, thereby advancing theoretical precision in the study of AI-mediated language learning.

The broader AI-related literature—including meta-analyses such as Xu et al. (2025), systematic reviews (Chalmers et al., 2021; Gkintoni et al., 2025; Goh and Aryadoust, 2025; Zhang and Aubrey, 2024), and theoretical frameworks (Sana and Fenesi, 2025; Xu, 2025)—is used to contextualize these 31 core empirical studies but is not counted in the primary empirical corpus. Additional empirical work examining working memory-relevant constructs in non-AI or AI-adjacent contexts, such as Teng (2024) and Xiao et al. (2025), further informs the interpretive framework.

By comparing the 27-study Interactive Era corpus with the 4-study AI-Mediated cluster and situating both within this broader context, the review seeks to identify continuities (what remains true across eras), transformations (what has fundamentally changed), and persistent measurement gaps (what researchers still cannot determine) in how working memory is conceptualized, engaged, and empirically assessed as language education transitions into the AI era.

This systematic review is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 Statement (Page et al., 2021). Figure 1 provides the PRISMA 2020 flow diagram documenting identification, screening, eligibility, and inclusion across both Phase 1 (Interactive Corpus) and Phase 2 (AI-Mediated Cluster), including categorized reasons for full-text exclusion with exact counts. The completed PRISMA 2020 checklist with page/section cross-references is provided in Appendix A (PRISMA 2020 Checklist). Because the review is explicitly historical–comparative and constructs two analytically distinct corpora (Interactive vs. AI-mediated) rather than estimating a single pooled effect, synthesis is conducted within each corpus prior to cross-era comparison, and contextual (non-primary) literature used for interpretation is reported separately from the primary empirical study count.

This study adopts a two-phase systematic review design with an explicit historical-comparative structure. The goal is not to collapse all studies into a single pooled estimate but to construct two analytically distinct yet comparable corpora, each analyzed in relation to the three guiding research questions (RQ1: design guidelines; RQ2: WM × AI affordances interactions; RQ3: boundary conditions and unintended consequences): an Interactive Corpus of 27 empirical studies examining working memory in traditional, non-AI technology-enhanced language learning environments from 2010 to 2024, and an AI-Mediated Cluster of four empirical studies from 2024 to 2025 providing detailed evidence on cognitive load and working memory-relevant processes in AI-mediated English as a foreign language and second language learning. These 31 studies constitute the primary empirical dataset. Additional meta-analytic, systematic review, theoretical, and AI-adjacent empirical work is drawn upon to interpret patterns and highlight gaps but is not included in the numerical count of the primary empirical corpus.

Phase 1 focused on the pre-AI Interactive Era. A systematic search was conducted for studies published between 2010 and 2024 that investigated working memory in interactive, technology-enhanced language learning contexts. Searches were conducted in major databases including Web of Science, Scopus, and ERIC. The core search string combined working memory with generic interactive technology terms: “working memory” AND (“computer-assisted language learning” OR “interactive language learning” OR “online language learning”). This string was adapted to each database’s syntax while preserving its conceptual structure. Limits were set to peer-reviewed journal articles, English language, and second or foreign language learning contexts. Reference lists of eligible papers and key theoretical and review articles were screened through backward citation tracking to identify additional studies that met inclusion criteria but might not have been captured due to terminology variation.

In total, 372 records were initially identified through database searching and other sources (Figure 1). Titles and abstracts were screened for relevance, leading to the exclusion of 162 records that clearly did not concern working memory or technology-enhanced language learning. The remaining 196 full-text articles were assessed for eligibility against predefined inclusion and exclusion criteria. After this full-text screening, 169 articles were excluded (categorized reasons and counts are reported in Figure 1).

Studies were included in the Interactive Corpus if they reported empirical data (experimental, quasi-experimental, correlational, or mixed-methods) on language learning or performance in interactive, technology-enhanced environments such as computer-assisted language learning software, multimedia platforms, hypermedia, non-AI captioned video, or online courses. Studies were also required to explicitly discuss working memory, either conceptually by framing results in terms of working memory demands or capacity, or empirically by using working memory measures as predictors, moderators, or covariates in relation to language learning outcomes. Additionally, studies needed to provide sufficient methodological detail to allow assessment of sample characteristics, instructional context, technological environment, and outcome measures. Studies were excluded if they focused exclusively on non-interactive or purely traditional face-to-face instruction without technology, did not mention working memory or working memory-related constructs, were purely theoretical or descriptive without empirical data, or implemented AI-mediated instruction, as these were reserved for Phase 2 screening.

Applying these criteria yielded 27 empirical studies constituting the Interactive Corpus. These studies all predate the widespread deployment of generative AI and biometrically adaptive systems in language education and do not involve AI-mediated instruction. Technologies examined include conventional multimedia computer-assisted language learning, hypermedia environments, captioned video without AI, online learning platforms, language games, and traditional data-driven learning using concordancers.

Phase 2 targeted the emerging AI-Mediated Era, focusing on identifying primary empirical interventions that both implemented AI-mediated language learning and provided rich cognitive load or working memory-relevant data suitable for comparison with the Interactive Corpus. To capture the rapidly emerging AI landscape, the search strategy was expanded to include AI-specific terminology alongside working memory-related constructs. Searches were conducted through 2025 in the same core databases, using a representative string such as: (“working memory” OR “cognitive load”) AND (“artificial intelligence” OR “AI-assisted” OR “generative AI” OR “chatbot” OR “intelligent tutoring” OR “adaptive learning” OR “AI-mediated” OR “virtual reality language learning”). Additional targeted searches used combinations such as “cognitive load” AND “AI-assisted writing,” “biometric feedback” AND “language learning,” “VR” OR “AR” AND “AI” AND “vocabulary learning,” “LLM” AND “second language acquisition,” and “working memory training” AND “L2.” These searches were supplemented by forward and backward citation tracking from key AI-in-education and AI-in-second language acquisition publications, as well as manual scanning of recent issues of high-relevance journals including Computer Assisted Language Learning, Language Learning & Technology, ReCALL, and Frontiers in Psychology. Identification, screening, eligibility, and inclusion counts for Phase 2 are reported in Figure 1 to document the full two-phase selection process in PRISMA 2020 format.

To retain a tight, analytically coherent cluster for detailed comparative analysis, Phase 2 applied more restrictive criteria. Studies were included in the AI-Mediated Cluster if they implemented AI-mediated language learning interventions such as generative AI-assisted writing, AI-orchestrated virtual and augmented reality instruction, biometric-adaptive AI reading platforms, or AI-mediated data-driven learning. Studies were also required to report quantitative data on cognitive load and/or working memory-relevant processing, including validated cognitive load scales, detailed subscale analyses, or process indicators such as eye-tracking interpreted in terms of working memory demands. Additionally, studies needed to involve English as a foreign language or second language learners in classroom-like or ecologically valid learning settings, and to provide sufficient statistical detail including means, standard deviations, effect sizes, F-statistics, partial eta squared values, and beta coefficients to support comparative analysis. Studies were excluded if they were systematic reviews, meta-analyses, or theoretical papers without primary empirical data, as these were used for contextual interpretation only. Studies were also excluded if they did not include cognitive load or working memory-relevant measures or examined AI in non-language-learning contexts. Based on these criteria, four primary empirical studies were retained as the AI-Mediated Cluster (Figure 1).

Yu (2025) conducted a pretest-posttest randomized controlled trial with 383 Chinese English as a foreign language learners examining AI-driven multimodal vocabulary instruction using virtual and augmented reality technologies including Unity, Oculus Quest 2, and Google MediaPipe combined with a generative AI chatbot based on ChatGPT-4. The study provides quantitative data on sensory overload boundaries and establishes an inverted-U relationship between channel complexity and retention. Yuan (2025) conducted a 12-week randomized controlled trial with 300 Chinese English as a foreign language learners investigating an AI-enhanced biometric-adaptive reading platform called Smart Sparrow that integrates eye-tracking using Tobii Pro X3-120 and physiological indicators to dynamically adjust text difficulty. The study demonstrates preemptive cognitive regulation and provides detailed cognitive load data from the Differentiated Cognitive Load Questionnaire subscales alongside eye-tracking indicators. Fan and Yao (2025) conducted a scale development and validation study with 546 participants across exploratory and confirmatory factor analysis samples for the Cognitive Load Scale for AI-Assisted L2 Writing, examining Chinese English as a foreign language learners using a generative AI chatbot called DeepSeek V3.1 for argumentative writing. The study provides a detailed four-factor profile of cognitive load redistribution across Authorial Core Processing, Critical Evaluation, Prompt Management, and Integrative Synthesis. Esfandiari and Allaf-Akbary (2024) conducted a three-arm randomized controlled trial with 93 Iranian advanced English as a foreign language learners comparing AI-mediated intelligent data-driven learning using Microsoft Copilot based on GPT-4, traditional corpus-based data-driven learning using AntConc, and conventional instruction for developing interactional metadiscourse markers. The study demonstrates very large effects for AI-mediated data-driven learning and theorizes working memory support via Schmidt’s noticing hypothesis.

To provide interpretive context for the 31-study empirical corpus, additional sources were systematically reviewed but not included in the numerical study count. These include meta-analytic work such as Xu et al. (2025), who synthesized 15 AI-assisted second language learning studies with 2,156 participants and reported large overall effects and potential developmental differences consistent with working memory compensation. Systematic and narrative reviews documenting the working memory measurement gap and AI’s cognitive implications were also consulted, including Chalmers et al. (2021), Zhang and Aubrey (2024), Goh and Aryadoust (2025), and Gkintoni et al. (2025). Theoretical and conceptual papers elaborating aptitude-treatment interaction frameworks and AI-related aptitude perspectives were reviewed, including Sana and Fenesi (2025) and Xu (2025). Empirical AI-adjacent studies with working memory-relevant data that inform but are not central to the two main corpora were also considered, including Teng (2024) and Xiao et al. (2025). These sources inform the Discussion and Future Directions sections, particularly the articulation of the measurement paradox and the argument for aptitude-treatment interaction-based AI research, but they do not alter the primary empirical synthesis count of 31 studies.

For each of the 31 empirical studies comprising 27 Interactive and 4 AI-Mediated, a standardized data-extraction template was used to capture bibliographic information including authors, year, journal, and country or region. Participant characteristics were recorded including sample size, age and educational level, first language, target language, and proficiency indicators. Instructional context was documented including course type, skills targeted, and duration and intensity of the intervention. Technology characteristics were captured including type of platform or tool, whether non-AI or AI-mediated, and presence of multimodality, adaptivity, or biometric integration. Working memory-related constructs and measures were recorded including direct working memory tests such as span tasks and n-back, cognitive load scales such as NASA-TLX, the Differentiated Cognitive Load Questionnaire, and the Cognitive Load Scale for AI-Assisted L2 Writing, and process indicators such as eye-tracking and physiological measures interpreted in terms of working memory demands. Study design was documented including experimental, quasi-experimental, correlational, scale development and validation, or mixed methods. Outcome measures were recorded including language performance such as vocabulary, reading comprehension, and writing quality, process indicators, and affective outcomes. Key findings were extracted specifically those related to working memory challenges such as overload and inequitable effects by working memory level and working memory affordances such as strategic use of working memory, dynamic regulation, and compensation.

Analysis proceeded in two stages. In the within-era synthesis stage, for the Interactive Corpus, themes were identified concerning how working memory was conceptualized as constraint, predictor, or moderator, how it was measured directly versus inferred, and how interactive design features affected working memory-related outcomes in terms of affordances versus overload. For the AI-Mediated Cluster, the focus was on how AI systems redistributed cognitive load across working memory subsystems, how they regulated working memory demands through reactive versus preemptive mechanisms, and what specific boundary conditions emerged such as channel limits in multimodal virtual and augmented reality instruction.

In the historical-comparative integration stage, findings from the two corpora were compared along dimensions defined by the conceptual framework, including working memory as constraint versus working memory as target for compensation and regulation, load reduction versus load redistribution, reactive adaptation versus preemptive regulation via biometrics, direct working memory measurement versus reliance on cognitive load proxies and process indicators, and general notions of overload risk versus empirically specified boundary conditions such as three to four concurrent channels in virtual and augmented reality instruction. This comparative structure underpins the Results, Discussion, and Implications sections, where the evolution of English as a foreign language education from static interactive environments to AI-mediated systems is analyzed in terms of continuities, transformations, and persistent empirical gaps regarding working memory.

The quality of the selected studies was assessed using the Mixed Methods Appraisal Tool (Pluye et al., 2011), which is designed to evaluate qualitative, quantitative, and mixed-methods research. The Mixed Methods Appraisal Tool was used to appraise study design, data collection, analysis, and reporting. For randomized controlled trials within both corpora, additional evaluation employed the Cochrane Risk of Bias tool (Higgins et al., 2021). For non-randomized quantitative studies, the Risk of Bias in Non-randomized Studies of Interventions tool (Sterne et al., 2016) guided assessment of confounding, selection, measurement, and reporting biases.

Notable quality considerations, especially pertinent to the AI-Mediated Cluster and contextual meta-analytic work, include evidence of publication bias favoring positive AI effects, as indicated by funnel plot asymmetry in the meta-analysis by Xu et al. (2025). The concentration of AI-mediated primary studies in EFL contexts, specifically Chinese and Iranian learners, limits generalizability to other languages, regions, and educational systems. The persistent absence of direct working memory measurement in AI-mediated intervention studies, despite heavy reliance on working memory-related theorizing, represents a significant methodological limitation. Quality assessment was conducted independently by two reviewers, with discrepancies resolved through discussion and consensus (Table 2).

These methodological foundations support the subsequent comparative analysis of how working memory is challenged, supported, and in AI contexts dynamically reconfigured across the two technological eras.

The review synthesized 31 primary empirical studies examining the challenges and affordances of working memory (WM) in technology-enhanced language learning. Twenty-seven studies formed the Interactive Corpus, representing traditional computer-assisted language learning (CALL) software, multimedia platforms, hypermedia environments, captioned video, and online learning systems published between 2010 and 2024. Four studies formed the AI-Mediated Cluster, representing generative AI chatbots, biometric-adaptive reading platforms, and VR–AR systems with large language model (LLM) integration published between 2024 and 2025. Across the corpus, research designs included randomized controlled trials, quasi-experimental studies, correlational investigations, scale development and validation work, and mixed-methods designs, conducted in classrooms, online platforms, language laboratories, and immersive virtual environments with learners of varied ages, proficiency levels, and first-language backgrounds.

The results are organized to foreground both the established patterns from traditional interactive environments and the emerging patterns from AI-mediated contexts. For each theme, findings from the 27 Interactive Corpus studies are presented first, followed by examination of how these patterns are transformed, extended, or problematized in the AI-Mediated Cluster and related recent reviews. Table 3 provides a study-level comparative evidence map of all included primary empirical studies (n = 31) and serves as the traceable basis for the cross-era synthesis reported in the Results section.

This table evidence-maps all included primary empirical studies using consistent coding rules to support transparent cross-era comparison. “WM measurement approach” is coded as Direct (psychometric WM task), Indirect (proxy indicators such as cognitive load scales and/or process measures interpreted as WM demand), or None (WM invoked conceptually without an empirical WM/CL operationalization). WM = working memory; AI = artificial intelligence.

Returning to the three research questions, this comparative synthesis illuminates a fundamental shift in the architecture of technology-enhanced language learning—from a hermeneutics of accommodation to a hermeneutics of co-regulation. Regarding RQ1 (design guidelines), the Interactive Corpus demonstrates that multimodal coordination, adaptive pacing, and schema activation support WM efficiency, whereas information density, technical disruptions, and unscaffolded complexity overload capacity. Regarding RQ2 (WM × AI affordances interactions), AI-mediated environments exhibit qualitatively distinct interaction patterns: rather than simply reducing load, they redistribute cognitive demands from encoding (phonological loop) to evaluation and integration (central executive), as evidenced by the CL-AI-L2W factor structure and biometric-adaptive process data. Regarding RQ3 (boundary conditions and unintended consequences), the review identifies empirical channel limits (3–4 concurrent sources in VR–AR), the measurement paradox (increasing theoretical invocation but decreasing direct assessment of WM), and the risk that load redistribution may disadvantage learners lacking metacognitive strategies, as well as potential over-scaffolding (i.e., overprotection) with downstream effects on metacognitive monitoring and executive-skill development.

In the pre-AI corpus, working memory was treated as a finite, static buffer: an invariant biological constraint around which instructional design had to be carefully engineered. Technologies were expected to adapt; cognition was assumed to be fixed. The design mandate was therefore to manage information density, sequencing, and modality so that task demands remained within known working-memory limits, consistent with the core tenets of Cognitive Load Theory and multimedia learning frameworks (Mayer, 2014; Sweller, 2020).

The emergence of AI-mediated environments, particularly those integrating generative AI and biometric sensing, destabilizes this arrangement. The unit of analysis is no longer the solitary learner but a dynamically coupled human–AI system. In this regulatory model, working-memory demands are not simply “respected” but actively monitored, redistributed, and, at least in principle, shaped over time. Biometric-adaptive systems that adjust text complexity in real time based on physiological markers of arousal and effort exemplify this shift from reactive accommodation to preemptive regulation (Yuan, 2025). Meta-analytic evidence that K–12 learners, whose working memory is still developing, can achieve gains comparable to or larger than adults when supported by AI pacing and feedback suggests that AI may function as a compensatory prosthesis rather than merely a gentler delivery channel (Xu et al., 2025). In this sense, AI is not simply a scaffold resting on a fixed cognitive foundation; it is an active participant in the regulation of task demands and cognitive resources.

A central theme emerging from this comparison is the move from load reduction to load redistribution. Traditional interactive environments, informed by Cognitive Load Theory, sought to lower overall cognitive load by minimizing extraneous demands and managing intrinsic complexity. AI-mediated environments, by contrast, alter the structure of the load rather than its sheer volume. Generative AI in L2 writing offloads lower-level encoding demands—such as lexical retrieval and syntactic assembly—that classically tax the phonological loop, while substantially increasing demands on central-executive processes: prompt formulation, output verification, and integrative synthesis of human and machine contributions (Fan and Yao, 2025). The factor structure and mean profiles of the CL-AI-L2W scale, with relatively lower load on authorial core processing and higher load on evaluation and management, empirically capture this redistribution. Similarly, AI-mediated data-driven learning automates manual pattern search and concordancing, freeing cognitive resources for higher-order noticing and internalization, as reflected in the substantial performance advantages of AI-supported DDL over traditional corpus-based tasks (Esfandiari and Allaf-Akbary, 2024).

This redistribution is double-edged. For learners with robust executive control and metacognitive strategies, AI may indeed create a cognitive surplus—that is, freed working memory capacity available for higher-order processing when routine demands are offloaded (Sweller et al., 2019)—by removing routine encoding burdens. For learners with fragile central-executive resources or limited experience in critical evaluation, the same tools may intensify perceived difficulty: the effort saved in generating language is immediately reallocated to monitoring, checking, and reconciling AI output with task requirements and personal goals. The benefits of AI are therefore conditional, not automatic, and hinge on learners’ preparedness for the new forms of cognitive work that AI-mediated tasks demand.

The review also exposes a deeper epistemic problem: the Measurement Paradox. As the field’s theoretical discourse on AI and cognition grows more sophisticated—foregrounding working memory, cognitive load, and aptitude–treatment interactions—the direct measurement of working memory has receded. Many of the pre-AI studies included validated span tasks or closely related measures, enabling explicit tests of how working-memory capacity constrained or enabled learning. In contrast, the AI-mediated intervention studies, despite their claims about compensation and optimization, rely almost exclusively on subjective cognitive load ratings, proficiency scores, or process indicators as proxies for working memory. Across multiple systematic reviews, working memory is repeatedly invoked as central to aptitude and AI-mediated learning, yet rarely measured with psychometric rigor (Chalmers et al., 2021; Goh and Aryadoust, 2025; Sana and Fenesi, 2025; Zhang and Aubrey, 2024; Xu, 2025). Teng’s (2024) finding that working memory influences vocabulary learning indirectly via proficiency illustrates the complexity of these relationships and underscores that nuanced, multi-step pathways cannot be inferred from load ratings alone.

This reliance on proxies creates a causal black box. When an AI intervention is associated with lower self-reported load and better outcomes, we cannot determine whether the AI simplified the task environment, improved learners’ allocation of existing working-memory resources, or contributed to durable changes in capacity. These mechanisms are conceptually distinct and carry very different implications for design, equity, and long-term development, yet they are empirically conflated when working memory is left unmeasured.

Domain-specific and neurocognitive evidence in the corpus further complicates the picture but also opens new possibilities. Findings that verbal working-memory measures, particularly backward digit span, differentiate high- and low-achieving L2 learners more strongly than spatial measures suggest that specific subsystems of working memory are especially critical for language learning (Xiao et al., 2025). Reviews of multisensory working-memory training using linguistically rich materials indicate that aspects of working memory may be more plastic than trait models assumed, with gains in capacity and neural efficiency accompanied by transfer to new auditory tasks (Gkintoni et al., 2025). Although these studies are not themselves AI-mediated, they point toward an underexplored design space in which AI could be used not only to compensate for working-memory constraints but also to deliberately train key verbal working-memory processes in ecologically valid language tasks.

Taken together, the present review suggests that the field stands at a crossroads. On one path lies a sophisticated narrative of AI as a co-regulator of cognition that compensates for developmental and individual differences in working memory. On the other lies a persistent methodological reticence to measure working memory directly, leaving fundamental causal questions unanswered. At present, the evidence base is insufficient to adjudicate longer-run developmental trade-offs; a balanced account therefore must specify where AI-mediated support may backfire through over-scaffolding (i.e., overprotection), observable as dependency on AI support, reduced independent strategy use, and reduced transfer under reduced-support conditions, with potential downstream costs for metacognitive monitoring and executive-skill development. These boundary conditions can be tested empirically by incorporating scaffold-fading schedules, explicit reduced-support phases, delayed transfer tasks, and metacognitive/executive outcome measures (e.g., monitoring accuracy, strategy-use indices, and executive-control measures) alongside immediate performance outcomes. Bridging this gap is essential if AI is to be understood—and responsibly deployed—not merely as a convenience layer on existing pedagogies, but as a transformative partner in the orchestration of cognitive effort.

This review has traced the evolution of working memory as a central construct in technology-enhanced language learning across two technological epochs. In the pre-AI era, working memory functioned as a hard constraint: an internal bottleneck around which well-designed systems attempted to navigate. In the emerging AI-mediated era, it becomes part of a distributed, dynamically regulated system in which cognitive demands can be sensed, reallocated, and, potentially, reshaped through continuous interaction with intelligent tools.

The analysis suggests that AI’s most consequential contribution is not simply its capacity to deliver content more efficiently, but its ability to reorganize the division of cognitive labor between human and machine. By offloading some processes and intensifying others, AI forces a reconceptualization of what it means to design, teach, and learn in ways that are sensitive to working-memory limits and possibilities. At the same time, the methodological apparatus of the field has not yet caught up with these conceptual advances. Without systematic, direct measurement of working memory and rigorous tests of how it interacts with AI affordances, claims about compensation, optimization, and training will remain aspirational.

The challenge, then, is twofold. Conceptually, the field must embrace a view of learning in which working memory is neither a fixed liability to be accommodated nor a simple trait to be correlated, but a dynamic resource that co-evolves with the tools through which learners engage with language. Methodologically, it must commit to an empirical program capable of illuminating the hidden dynamics of this human–AI collaboration. Only by holding these two commitments together can AI-mediated language learning move from speculative promise to evidence-based practice—and, in doing so, transform working memory from a limiting boundary into a central site of educational design and innovation.