Effects of exercise interventions on memory in depression: a three-level meta-analysis

Inclusion criteria

Participants were patients with depression who met either the diagnostic criteria of the International Classification of Disease (ICD) or the criteria of the Diagnostic and Statistical Manual of Mental Disorders (DSM) and had no other mental diseases. The intervention for the experimental group consisted of either exercise alone or exercise combined with the components of the control group intervention. The minimum duration of the exercise intervention was set at three weeks, which aligns with the average hospitalization period (approximately three to four weeks) for patients receiving inpatient depression treatment. This threshold was chosen for pragmatic reasons, reflecting typical clinical practice rather than being empirically determined. The intervention contents of the control group were irregular exercise or conventional treatment, including drug therapy, occupational therapy, stretching exercise, health education, placebo, etc; The outcome index was memory memory-related index; The type of study was a randomized controlled trial.

Exclusion criteria

The subjects were non-depressed patients. The languages used in the study were not English or Chinese. Articles in which data could not be obtained or extracted for estimating effect size even after contacting the authors. Studies using a single bout of exercise as an intervention. Studies using animal models, conference abstracts, book chapters or reviews.

Literature screening

Two researchers (YN Z and XL Z) independently screened the literature according to the inclusion and exclusion criteria. First, the retrieved literature was imported into Endnote X9 to eliminate duplicate literature and read the titles and abstracts of literature for preliminary screening. Secondly, the full text was screened again to ensure that the content of the study and the outcome indicators met the inclusion criteria. Then, the final included literature was determined. If two researchers had disagreements, a third researcher (C L) joined in the discussion to make the final decision.

Data extraction and coding strategy

Two authors (YN Z and XL Z) extracted data separately by using a pre-developed extraction form in Microsoft Excel. The agreement of extraction between the two was 96.82%. The extracted information included basic information (author, year of publication, country, age, sample size), experimental characteristics (two groups of intervention content and exercise intensity), and all memory-related outcome indicators in each literature. The outcome data were collected during pre-intervention, during-intervention and post-intervention. If the data was missing or unclear, the original author would be contacted through email; when the information extracted by two researchers (YN Z and XL Z) was inconsistent, a third researcher (C L) would join in the discussion and make a decision together.

Due to the lack of commonality in exercise frequency, we could not perform subgroup analyses based on this variable; therefore, we coded only exercise type, duration, intensity, intervention content of the experimental group, and age. We categorized exercise type as aerobic, strength, and mind-body exercise; coded exercise time as short (≤60 min) and long (>60 min); classified exercise duration as ≤12 weeks and >12 weeks; and defined exercise intensity as low-to-moderate, moderate, and moderate-to-vigorous. (Ren et al., 2023; Ren et al., 2024). Additionally, we coded the intervention content of the experimental group as either exercise alone or exercise combined with other therapies. Age was categorized into young adults (18–44 years), middle-aged adults (45–64 years), and older adults (≥ 65 years).

Assessment of study quality

Two reviewers (YN Z and XL Z) independently assessed the risk of bias in the included studies using the Cochrane Risk of Bias 2 (RoB 2) tool. RoB 2 was selected instead of the PEDro scale because it represents the current Cochrane-recommended standard for randomized trials and is more appropriate for exercise interventions with psychological or cognitive outcomes, where items such as participant or therapist blinding in PEDro are often infeasible and may artificially deflate quality scores. This modification was also made in response to prior reviewer feedback, and the review team agreed that RoB 2 would provide a more rigorous and transparent assessment framework. RoB 2 comprises five domains: randomization process; deviations from intended interventions; missing outcome data; measurement of the outcome; and selection of the reported result (Sterne et al., 2019). Each domain is rated as low risk of bias, some concerns, or high risk of bias. An overall judgment of low risk of bias is assigned when all domains are rated low risk; some concerns is assigned when at least one domain is rated some concerns and none is rated high risk; and high risk of bias is assigned if any single domain is rated high risk (Sterne et al., 2019; Wang et al., 2025). In case of disagreement, a third researcher (C L) would discuss and make the final decision.

GRADEpro software was used to evaluate the quality of outcome evidence. From “high risk of bias in study design or implementation”, “high heterogeneity or inconsistency of results”, “indirectness of evidence”, “imprecision of results” and “high probability of publication bias” were used to evaluate downgrade or upgrade. Ultimately, each level of evidence was rated as “high”, “moderate”, “low” and “very low”(Li et al., 2023).

Statistical analysis

This study utilized the metafor package in R version 4.3.0 to conduct a three-level meta-analysis using a random effects model (Assink & Wibbelink, 2016; Cui et al., 2024; Viechtbauer, 2010). This model considers the following three categories of effect size (ES) variability: sampling variance (level 1); within-study variance (level 2); and between-study variance (level 3) (Cheung, 2014). With the use of the restricted maximum-likelihood estimation (REML) method, pooled ES of exercise on memory was determined using Hedges’ g and 95% confidence intervals (CI). When the ES was divided according to Hedges’ g, P < 0.05 was statistically significant. Hedges’ g<0.20 refers to a small effect, 0.20–0.49 is a small-to-moderate effect, 0.50–0.79 is a medium effect, and ≥ 0.80 is a large effect (Cohen, 1992). 95% prediction intervals (PI) were calculated to determine the expected range in which an ES in future identical studies will fall.

Testing for statistical significance in level 2 (within-study) and level 3 (between-study) variances were performed using likelihood ratio tests (LRT).The I² statistic was used to estimate of total between-study heterogeneity (1–49% indicating low, 50–74% indicating moderate, and 75–100% indicating high) (Higgins et al., 2003). To evaluate the robustness of the findings, we conducted a series of additional sensitivity analyses following recommended meta-analytic practices. Publication bias was assessed using Egger’s regression test and a contour-enhanced funnel plot (Egger et al., 1997). Influence diagnostics (including studentized residuals, Cook’s distances, DFFITS, covariance ratios, hat values, and QE deletion diagnostics) were computed to identify potential influential or outlying effect sizes (Viechtbauer & Cheung, 2010). Leave-one-out analyses were performed to test the stability of the pooled effect. Based on prior evidence, exercise intensity was designated a priori as the primary moderator (Liu et al., 2024). The other moderators (duration, intervention content, age, intervention type, session time) were considered exploratory. For the exploratory tests, we controlled multiplicity using the Benjamini–Hochberg false discovery rate (FDR, q = 0.05) and report both unadjusted omnibus P and FDR-adjusted q values (Benjamini & Hochberg, 1995).

Literature search results

Seven databases yielded 3,648 records; citation chasing added four (total 3,652). EndNote X9 de-duplication removed 103, leaving 3,549 for title/abstract screening. Of 52 full texts sought, 11 were unobtainable; 41 were assessed, and 25 were excluded (outcome mismatch, five; review/abstract, five; intervention ineligible, three; non-randomized controlled trial (RCT), three; data not extractable, seven; other, (2). Sixteen studies remained (14 English, two Chinese; including one grey source) (see Fig. 1).

Characteristics of the included literature

The 16 studies were published between 2001 and 2022, originating from Germany (two), the USA (four), Switzerland (one), Denmark (three), India (two), and China (four). The age of patients with depression ranged from 31.50 to 70.55 years. The intervention contents of the experimental group were aerobic exercise, strength training, Taichi, aerobic exercise combined with conventional rehabilitation, and yoga combined with conventional rehabilitation, whereas the intervention contents of the control group were occupational or artistic therapy, placebo, etc; Exercise intensity was divided into low-to-moderate intensity, medium intensity and medium-to-vigorous intensity. The outcome indicators were DSTB(Digit span Test backward), VIMTLT (Verbaler Lern-und Merkfähigkeitstest learning, total) and so on, as shown in Table 2.

Risk of bias in the included studies

Risk of bias was assessed for the included studies using the RoB 2 tool. Across the five domains, 15 studies were judged low risk for the randomization process; eight were low risk for deviations from intended interventions; 14 were low risk for missing outcome data; 13 were low risk for measurement of the outcome; and six were low risk for selection of the reported result. Overall, one study was rated high risk of bias, two were low risk of bias, and 13 raised some concerns (Fig. 2).

The interventional effects of exercise on memory in patients with depression

We employed random-effects models to aggregate effect sizes. For verbal memory, the three-level random-effects model showed a small, statistically significant effect of exercise (g = 0.17, 95% CI [0.02–0.32], p = 0.03), with a 95% PI of −0.06 to 0.40, as shown in Fig. 3. The two-level random-effects model yielded a similar result (g = 0.17, 95% CI [0.03–0.32], p = 0.02), with a 95% PI of 0.03 to 0.32, indicating a likely minimal in magnitude but consistent benefit. Heterogeneity was low (Q = 13.59, p > 0.05).

For visual memory, the three-level random-effects model indicated no statistically significant effect (g = 0.27, 95% CI [−0.00 to −0.54], p = 0.05), with a PI spanning the null (−0.00 to 0.54) (Fig. 4). The two-level random-effects model produced nearly identical estimates (g = 0.27, 95% CI [−0.00–0.54], p = 0.05), with a PI spanning the null (−0.00 to 0.54), reinforcing the conclusion that any effect, if present, is likely minimal and uncertain. Heterogeneity was again low (Q = 3.37, p > 0.05).

Influence analysis

We conducted formal influence diagnostics (studentized residuals, Cook’s distance, hat values, and DFBETAs) and complemented them with a leave-one-study-out analysis. For verbal memory, deleting any single study yielded pooled estimates of g = 0.16–0.19 with p = 0.02–0.04, leaving the direction and statistical significance unchanged. For visual memory, the pooled estimate ranged g = 0.24–0.30 with p = 0.04–0.08; although the direction remained positive, the p-values straddled 0.05, indicating marginal and non-robust evidence. Influence diagnostics did not identify any outliers affecting either outcome. This pattern is illustrated in Figs. 5 and 6 and is further supported by the data presented in Tables S1 and S2.

Heterogeneity analysis

In the analysis of verbal memory, the log-likelihood ratio test indicated that the between-study variance (level 3) was not significant, χ² = 0.17, p = 0.34, suggesting that the three-level model did not outperform the two-level model. The within-study variance (level 2) was also not significant, χ² = 0.00, p = 0.50, further indicating that the three-level model did not provide a superior fit compared to the two-level model. Despite this, employing the three-level model remains meaningful. In terms of overall variance, the within-study variance (level 2) accounted for 8.60%, and the between-study variance (level 3) accounted for 2.87%.

In the analysis of visual memory, the log-likelihood ratio test revealed that the between-study variance (level 3) was not significant, χ² = 0.00, p = 0.50, indicating that the three-level model did not outperform the two-level model. The within-study variance (level 2) was also not significant, χ² = 0.00, p = 0.50, further suggesting that the three-level model did not provide a superior fit compared to the two-level model. Nevertheless, the use of the three-level model remains meaningful. In terms of overall variance, the within-study variance (level 2) accounted for 1.25%, and the between-study variance (level 3) accounted for 9.66%.

Although the three-level model did not significantly improve fit, we retained it because it reflects the hierarchical data-generating process with multiple outcomes nested within studies and appropriately accounts for within-study dependence; moreover, variance-component tests are underpowered when parameters lie on the boundary, so near-zero estimates alone are not sufficient grounds to collapse the hierarchy, and retaining the structure yields valid—and typically more conservative—inference (Van den Noortgate et al., 2013).

Moderator analysis

Since exercise did not affect visual memory in patients with depression, we examined moderators only for verbal memory. Exercise intensity was pre-specified as the primary moderator, whereas exercise duration, type, session time, intervention content of the experimental group, and age were treated as exploratory; for exploratory moderators, multiplicity was controlled using the Benjamini–Hochberg FDR, as shown in Table 3.

Exercise duration was not a statistically significant moderator (F = 3.04, p = 0.09, q = 0.24). Shorter interventions (≤12 weeks) showed a small, statistically significant effect (g = 0.27, 95% CI [0.09–0.46], p < 0.01), whereas longer interventions did not (g = 0.02, 95% CI [−0.21–0.25], p = 0.87). This suggests that any benefits may emerge relatively early and may not increase with longer treatment durations.

Exercise type was also not a significant moderator (F = 2.48, p = 0.10, q = 0.24). Neither aerobic exercise (g = 0.08, 95% CI [0.09–0.25], p = 0.36) nor strength training (g = 0.06, 95% CI [−0.59–0.72], p = 0.85) demonstrated significant effects. However, mind–body exercise yielded a small-to-moderate and statistically significant effect (g = 0.43, 95% CI [0.16–0.71], p < 0.01), indicating a potentially meaningful benefit for this exercise modality.

Session time was not a significant moderator (F = 0.25, p = 0.62, q = 0.62). Shorter sessions (≤60 min) produced a small but statistically significant effect (g = 0.18, 95% CI [0.02–0.35], p = 0.03), whereas longer sessions did not (g = 0.07, 95% CI [−0.38–0.51], p = 0.77). These findings suggest that briefer sessions may be more feasible and still beneficial.

Exercise intensity emerged as a significant moderator (F = 3.39, p = 0.04). Low-to-moderate intensity produced a small-to-moderate, statistically significant effect (g = 0.43, 95% CI [0.16–0.71], p < 0.01). Moderate intensity alone did not reach significance (g = 0.24, 95% CI [−0.05–0.53], p = 0.11), nor did moderate-to-vigorous exercise (g = 0.00, 95% CI [−0.20–0.20], p > 0.49). This pattern suggests that more strenuous activity may not yield additional cognitive benefits and could even reduce adherence or tolerability.

Intervention content (exercise alone vs. combined with other therapies) was not a significant moderator (F = 2.20, p = 0.14, q = 0.24). Exercise alone did not yield significant effects (g = 0.09, 95% CI [−0.09–0.27], p = 0.32). However, combined interventions produced a small but statistically significant effect (g = 0.31, 95% CI [0.09–0.54], p = 0.01), suggesting potential synergistic benefits when exercise is integrated with other therapeutic approaches.

Age group was not a statistically significant moderator (F = 0.87, p = 0.43, q = 0.53). Young adults demonstrated a small but statistically significant effect (g = 0.22, 95% CI [0.03–0.41], p = 0.03), whereas middle-aged adults (g = 0.03, 95% CI [−0.25–0.30], p = 0.85) and older adults (g = 0.32, 95% CI [−0.18–0.81], p = 0.19) did not show significant improvements. Given the limited number of studies involving older adults, these results should be interpreted cautiously.

Publication bias

We assessed publication bias using a funnel plot and Egger’s regression test. The results indicated that the funnel plot was largely symmetric. Additionally, Egger’s regression test suggested that there was no publication bias regarding the effects of exercise on verbal memory (t = 1.79, p = 0.08) and spatial memory (t = 0.59, p = 0.56) in patients with depression, as shown in Figs. 7 and 8.

GRADEpro evidence quality assessment

We conducted an evidence-quality assessment for verbal memory and visual memory. Due to the lack of distinction regarding the severity of depression among the included studies, the level of inconsistency was downgraded by one level, resulting in an overall evidence quality rating of moderate (see Table 4).

Main study findings

This study indicates that exercise can improve verbal memory in patients with depression; however, it does not demonstrate an effect on visual memory. Both findings are classified as moderate evidence, suggesting that future research may modify these conclusions. Although the pooled effect on verbal memory was small (g = 0.17) and the 95%PI included zero, such modest effects—particularly when tied to modifiable behavioral factors—may still be clinically relevant. A g-value of 0.17 roughly corresponds to a 4–6% improvement on standardized verbal memory tests, representing a subtle but potentially meaningful enhancement in everyday cognitive functioning.

The impact of exercise on memory in patients with depression

This study found that exercise can improve verbal memory in patients with depression, whereas no effect was observed for visual memory. Although the effect on verbal memory was small, between-study uncertainty suggests that this benefit may not consistently replicate. In contrast, Sun et al. (2018) conducted a traditional meta-analysis in which effect sizes from individual studies were pre-aggregated (five studies for verbal memory and three studies for visuospatial memory), yielding pooled effects of g = 0.09 (n = 577) for verbal memory and g = 0.35 (n = 302) for visuospatial memory; neither reached statistical significance. Similarly, Brondino et al. (2017) aggregated effect sizes from five studies on verbal memory and three studies on visuospatial memory and applied a conventional two-level meta-analytic model. Their findings indicated no improvement in verbal memory (g = 0.05, n = 577), but a small improvement in visuospatial memory (g = 0.24, n = 302). Unlike Sun et al. (2018), Brondino et al. (2017) further differentiated the effects of different exercise modalities. This discrepancy may be attributed to our use of a three-level meta-analysis method, which allowed for a comprehensive statistical analysis of all data related to verbal learning and memory as well as visual learning and memory. This approach minimized the issue of information loss that could reduce statistical power and mitigated the potential inflation of results due to correlations between effect sizes. Previous research employing three-level meta-analyses has yielded similar conclusions. Ren et al. (2024) also employed a three-level meta-analysis, incorporating eight studies, and found that aerobic exercise can improve memory in patients with depression. Unfortunately, they did not further differentiate between verbal memory and visual memory in their analysis. By explicitly modeling the dependency structure among multiple outcomes within studies, our three-level model reduced within-study redundancy, prevented disproportionate weighting of studies that reported more outcomes, and yielded more conservative and arguably more robust pooled estimates than would likely have been obtained using a traditional two-level approach with pre-aggregated effect sizes.

Taken together, these methodological features mean that the present synthesis contributes several elements that previous meta-analyses could not capture. First, it provides domain-specific pooled estimates for verbal and visual memory rather than a single global memory outcome, thereby clarifying that observable exercise-related benefits appear to be confined to verbal memory. Second, the three-level structure, combined with an expanded study base (including Chinese databases and gray literature), allowed a more systematic evaluation of potential moderators (e.g., intensity, session duration, exercise type, and intervention format), offering preliminary guidance on which exercise prescriptions may be most promising for targeting memory outcomes in depression.

From a mechanistic perspective, our findings align with current neurobiological frameworks linking exercise to domain-specific cognitive enhancement. Aerobic exercise promotes hippocampal neuroplasticity through increased BDNF expression, neurogenesis, and dentate gyrus remodeling-mechanisms known to preferentially support verbal/episodic memory(Dinoff et al., 2018; Erickson et al., 2011). In contrast, improvements in visual or visuospatial memory rely more heavily on prefrontal-dependent pathways, such as enhanced cerebral perfusion, synaptic plasticity, and HPA-axis regulation (Stillman et al., 2016). The absence of a significant effect on visual memory in the present study may therefore reflect weaker engagement of these prefrontal-dependent mechanisms under the exercise conditions typically implemented in the included trials. Exercise may thus exert stronger effects on hippocampal-dependent verbal memory relative to visual memory, providing a biologically plausible explanation for the observed dissociation (Kandola et al., 2019).

Characteristics of exercise

Due to the inability to categorize exercise frequency, this study focused exclusively on the effects of exercise type, session time, intensity, and total duration on verbal memory. Intensity (pre-specified as the primary moderator) showed a significant omnibus effect (F = 3.39, p = 0.04), with a small-to-moderate benefit for the low-to-moderate category (g = 0.43, 95% CI [0.16–0.71]); the other two intensity levels were not significant. Given the small magnitude and between-study uncertainty, this pattern should be interpreted cautiously. In contrast, the other two exercise intensity levels did not show improvements in verbal memory. Liu et al. (2024) also reached similar conclusions, finding that exercise intensity has a moderating effect on working memory in depressed patients, with low-intensity exercise being the most beneficial. However, other studies have reported different findings. Ren et al. (2023) and Ren et al. (2024) discovered that exercise intensity was not a moderating factor for improving executive function and cognitive abilities in depressed patients, suggesting that “moderate-to-vigorous” intensity produced the most favorable outcomes. These divergences likely reflect (i) differences in outcome domains—our analysis targets verbal memory specifically, whereas prior work examined executive or general cognition; (ii) different intensity categorizations (three levels here vs. two levels previously); (iii) analytic choices; and (iv) contextual features such as supervision, adherence, and tolerability. From a mechanistic standpoint, higher intensities may reduce affective valence and tolerability (Ekkekakis & Lind, 2006), potentially lowering adherence and diluting observable gains (Vandoni et al., 2016). Overall, the current evidence suggests a small, context-dependent advantage for verbal memory at low-to-moderate intensities, while alternative patterns reported elsewhere may stem from domain, categorization, and context differences rather than direct contradictions.

Exercise duration was not a moderating factor, and only “≤12 weeks” showed a small-to-moderate effect (g = 0.27). Ren et al. (2024) also posited that exercise for “≤12 weeks” can improve cognitive function in patients with depression. However, Liu et al. (2024) contended that exercise lasting 12–16 weeks is necessary to enhance working memory in depressed patients, while Ren et al. (2023) suggested that exercise for “≥13 weeks” is required to improve executive function. This discrepancy may be attributed to differences in outcome measures. Although our study indicated that exercise of “≤12 weeks” is effective, the studies we included had a minimum exercise duration of 3 weeks. Some research indicates that significant improvements in depressive symptoms often occur within the initial weeks of exercise participation (Zhao, 2014). Conversely, excessively prolonged exercise regimens may lead to boredom and frustration in patients with depression, resulting in increased dropout rates.

Session time was not a moderating factor, and only “≤60 minutes” demonstrated a small effect (g = 0.18). This finding is consistent with previous research. Ren et al. (2023) and Ren et al. (2024) indicated that “≤60 minutes” can enhance executive function and cognitive abilities in patients with depression. Prolonged exercise durations may lead to feelings of fatigue, dehydration, and disruptions in neuronal homeostasis, ultimately impairing cognitive function (Liu & Zeng, 2015).

Exercise type was not a moderating factor, with only “Mind–Body Exercise” demonstrating a small-to-moderate effect (g = 0.43). This result aligns closely with those of Liu et al. (2024) and Sun et al. (2018). Conversely, Ren et al. (2023) posited that only aerobic exercise can enhance executive function in patients with depression. We propose that “Mind–Mind-body exercise”, which includes practices such as yoga and tai chi, integrates physical postures, breath regulation, and relaxation techniques to achieve a dynamic balance between the body, mind, and external environment. This integration may be particularly effective in alleviating negative emotions (Yang et al., 2019). Moreover, these practices require attention and involve multiple cognitive processes, such as visual-spatial awareness and memory, to maintain stable body postures, thereby potentially enhancing perceptual functions and cognitive abilities, including memory (Zou et al., 2019).

From a clinical standpoint, the present findings suggest that low-to-moderate intensity exercise performed for ≤60 min per session over ≤12 weeks may offer small but potentially meaningful benefits for verbal memory in adults with depression. Mind–body modalities such as yoga or tai chi may be especially suitable for individuals with low motivation or limited tolerance for higher-intensity activity. These patterns provide preliminary guidance for clinicians designing exercise prescriptions tailored to patients’ cognitive needs and functional capacities.

Study design and sample characteristics

Subgroup analysis was performed on the intervention content of the experimental group and the age of the patients with depression. The intervention content was not identified as a moderating factor, with only “exercise combined with other therapies” demonstrating a small-to-moderate effect (g = 0.31). Previous studies have reported similar findings. Solely engaging in exercise did not improve the overall cognitive status of patients with depression, nor did it enhance specific cognitive dimensions such as attention and memory (Sun et al., 2018). Additionally, research indicates that combining antidepressant medication with appropriately intense aerobic exercise may more effectively ameliorate cognitive symptoms in patients with depression (Greer et al., 2015; Olson et al., 2017).

Age was not a moderating factor, with only “young adults (18–44 years)” demonstrating a small-to-moderate effect (g = 0.22). However, this finding is not consistent with previous research. Ren et al. (2023), Ren et al. (2024) indicated that exercise improved executive function and cognitive function only in “middle adults (45–59 years)” (Ren et al., 2023; Ren et al., 2024). Additionally, other studies have shown that exercise can enhance executive function in older adults aged 60 and above with mild cognitive impairment (Brandt et al., 2009). This discrepancy may be related to the limited number of studies involving participants over 65 years of age in the current research, as only two studies were included. Therefore, future investigations should focus more on the memory of elderly patients with depression. Furthermore, this may also relate to differences in outcome measures, as prior studies have placed greater emphasis on executive function and overall cognitive performance.

Strengths and limitations

This study has several advantages. It is the first to utilize a three-level meta-analysis to investigate the effects of exercise on verbal and visual memory in patients with depression, predicting a 95% CI, and it is the first to find no improvement in visual memory. Additionally, this research builds upon prior studies by expanding the literature search and including grey literature to mitigate publication bias, enhancing the reliability of the results. Furthermore, the study explores various factors affecting memory, providing evidence-based support for the development of exercise interventions targeting memory in depressed patients.

However, several limitations should be noted. Although the three-level approach addresses the methodological limitations of traditional meta-analyses that assume independence among effect sizes, the small number of included studies and uneven subgroup sizes limited its statistical advantages. All subgroup analyses should therefore be interpreted as exploratory. Additionally, restricting the search to Chinese- and English-language publications introduces potential language bias and may limit global generalizability.

Considerable heterogeneity also existed across studies in terms of exercise protocols (type, supervision, progression), participant characteristics (depression severity, medication use, comorbidities), and measurement tools. These factors likely contributed to residual between-study variability and may have obscured differential responses to exercise. Furthermore, despite inclusion of Chinese literature, non-Western samples remained under-represented, highlighting the need for broader geographical diversity in future work.

Finally, limitations in available data prevented stratification by depression severity or examination of additional effect modifiers such as baseline cognitive status. Although multiplicity was controlled using the Benjamini–Hochberg FDR, false-positive findings cannot be entirely ruled out. Future RCTs should implement standardized memory assessments, evaluate dose–response relationships, and incorporate mechanistic biomarkers (e.g., BDNF levels, HPA-axis indicators) to clarify pathways through which exercise influences memory.