When Memories Keep Grief Alive: A Dynamic Model of Memory Characteristics and Prolonged Grief

Abstract

Introduction

Prolonged grief disorder (PGD) is a condition marked by sustained separation distress. Although cognitive models suggest that memory processes are critical in PGD, the specific characteristics of loss-related memories have received limited longitudinal attention. This study examined how loss-related memory characteristics and prolonged grief symptoms (PGS) influence each other over the first 18 months of bereavement.

Method

A community sample of 275 bereaved adults (79% women, M_age = 46.4) completed assessments of PGS and the Oxford Grief-Memory Characteristics Scale at three time points (0–6, 6–12, and 12–18 months post-loss). Panel graphical vector autoregressive (GVAR) models decomposed contemporaneous and temporal within person associations over time.

Results

Memory characteristics emerged as the most influential variables over time. Memories high in re-experiencing qualities predicted increasing difficulty accepting the death and a greater range of memory triggers over time, which predicted subsequent difficulty moving forward. Predominantly negative memories were associated with subsequent heightened yearning, while visceral consequences of memories predicted later loneliness, and disconnection from memories of the past self-predicted role confusion. Re-experiencing qualities and the PGS emotional numbness had a bidirectional link. Contemporaneous networks showed dense connections among memory characteristics and symptoms within the same time point.

Conclusion

Overall, the findings demonstrate that loss-related memory processes are predictors of later change in PGS during the first 18-month post-bereavement. This highlights maladaptive memory characteristics as promising mechanistic targets for intervention and supports therapeutic approaches that directly address how memories of loss are encoded, retrieved, and interpreted.

Introduction

The death of a loved one is a stressful life event that most people adapt to over time. However, approximately one in 20 bereaved people develop a persistent and impairing form of grief [1], which has recently been introduced into the 11th edition of the International Classification of Diseases and related Health Problems (ICD-11) [2] and the text revision of the 5th edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5-TR) [3] as prolonged grief disorder (PGD). PGD is characterized by intense longing for a deceased or persistent preoccupation with the person, accompanied by additional symptoms such as sadness, identity disruption, or difficulty accepting the death. These symptoms persist for more than 6 months (ICD-11) or 12 months (DSM-5-TR) after the death and cause significant impairment.

The death of a close person affects the autobiographical memory, defined as the collection of memories of personal events from one’s own life [4]. Theoretical models of PGD highlight, among other factors, the role of memory processes [4–6] in the development and maintenance of the disorder and a range of memory processes have been proposed. First, prolonged grief symptoms (PGS) are thought to be partly driven by insufficient elaboration and integration of the loss into the autobiographical memory [6]. In adaptive grief, the reality of the death is thought to gradually become linked with memories of the person and the relationship and existing knowledge about the self and the future. In PGD, this process is thought to be disrupted which leaves memories of the deceased and the death event poorly connected to other autobiographical memories and knowledge [6]. Such poor loss memory integration has two consequences: easy triggering of unwanted memories of the deceased and the loss by cues in the environment and restricted access to other autobiographical memories when remembering the deceased and the loss [7]. The easy triggering of unwanted memories is also addressed by theories that posit that symptoms arise from the reactivation of associative memory networks when reminders are present [5, 7]. In response to experiencing easily triggered unwanted memories, patients use maladaptive control strategies and avoid internal and external reminders of the loss, which further prevents memory integration, thus leading to a maintenance cycle in which poorly integrated highly accessible memories continue to drive emotional distress and maladaptive behaviour. The poor access to autobiographical memories that do not involve the deceased is also predicted by the theory that, in PGD, self-identity is closely entwined with the deceased [4].

However, fine-grained longitudinal studies of memory processes after bereavement remain scarce, and the present study therefore examines how memory characteristics contribute to the development of PGS. Empirical studies have shown that more frequent memories and mental images related to the loss are associated with higher levels of PGS [8–10]. In particular, mental images related to the moment of death (e.g., an accident or of being present at the time of death) were associated with greater prolonged grief severity than other types of loss-related imagery [10]. Another study demonstrated that positive intrusive images of the deceased were associated with grief severity, but not with depression and anxiety, suggesting that experiencing positive memories as painful may be specific to grief [8]. This phenomenon was further described in qualitative work with bereaved participants with and without PGD: positive memories acquired distressing qualities because they highlighted the absence of the deceased and reinforced the awareness of the loss [11]. This difficulty may be further compounded by a broader retrieval bias. For example, Robinaugh and McNally [12] found that bereaved individuals with pathological grief were able to recall specific autobiographical memories involving the deceased but were significantly impaired at retrieving specific memories or imagining future events that did not include the deceased. This pattern suggests that in PGD, the loss dominates what is recalled from autobiographical memory, leaving the individual oriented to towards the deceased in memory and limiting access to positive non-loss memories, which may fuel yearning and inhibit re-organizing the identity after the loss [4].

Detailed empirical investigations of specific characteristics of unwanted loss-related memories remain limited compared to the more developed literature on posttraumatic stress disorder (PTSD). PGD and PTSD are distinct, yet related disorders that can develop in response to a stressful event (i.e., the loss of a close person in PGD and a traumatic event in PTSD). The dual representation theory of PTSD [13] posits different representations of traumatic events: inflexible, lower level sensory representations (S-reps), and flexible contextually bound representations. In PTSD, dysfunctional encoding is thought to lead to relatively stronger S-reps that are easy to trigger independently of the contextually bound representations, leading to unwanted sensory intrusions. Ehlers and Clark’s [7] cognitive model of PTSD specifies the encoding conditions that produce poorly integrated, disjointed memories and the nature of memory triggers: a predominance of perceptual processing at the expense of conceptual and self-referent processing during the worst moments of trauma is thought to result in memories that are poorly integrated into the autographical memories base and when triggered, lack context in time, place, and information that would update their threatening meaning [14]. High perceptual priming leads to easy triggering by perceptually similar cues. This has the consequence that when these unwanted memories are triggered, they appear to be happening in the present and elicit the same emotions and bodily sensations as during the trauma, while other autobiographical material is difficult to access. Given that theoretical accounts of PGD emphasize the integration of loss-related memories into autobiographical memory [6] and highlight the role of unwanted loss memories and their triggers [15], these accounts in PTSD can inform our understanding of loss-related memory characteristics.

Loss-Related Autobiographical Memory Processes

Smith and colleagues [16] investigated aspects of memory, informed by the PTSD and PGD literature reviewed above. The Oxford Grief-Memory Characteristics Scale (OG-M) [16] was developed to capture the features of loss-related memories and was based on in-depth interviews with cancer-bereaved individuals with and without PGD [11]. Although previous psychometric analysis of the OG-M indicated a unifactorial scale, conceptually the items map onto six theoretically informed clusters of loss memory characteristics and assess related but distinguishable aspects of loss memories.

The first cluster’s items assess the phenomenological qualities of unwanted memories of the loss, i.e., their frequency, distress, and “nowness” quality, summarized as re-experiencing qualities. In interviews, bereaved individuals described sensory reliving of moments surrounding the death, for example, feeling transported back to the smell and atmosphere of the hospital [11]. These loss memories feel emotionally immediate and “alive” and their features closely parallel intrusive memories in PTSD [7].

The other five clusters cover the triggers and consequences of memory activation (such as visceral responses, proximity-seeking impulses, and trigger sensitivity) and broader properties of how the loss influences memory recall (e.g., the predominance of negative content in recall, feeling disconnected from how the self is remembered). Importantly, the OG-M specifically targets both unwanted, intrusive memory phenomena as well as voluntarily recalled memories related to the loss and the deceased.

The second cluster captures that memories can entail visceral consequences that can manifest as bodily pain or fatigue triggered by recollections of the loss. One bereaved person described feeling the knots in the pit of their stomach in response to an intrusive image of their loved one in their hospital bed [11].

Third, loss-related memories can give rise to proximity-seeking impulses, including sudden urges to search for, comfort, or be close to the deceased [11]. These impulses are thought to be fuelled by intrusive memories because they maintain a sense that the deceased is only temporarily absent. They can also be understood within attachment-based models as reflecting a failure to update internal working models following the loss [17]. As a result, the deceased continues to be represented as physically and emotionally accessible, and memory activation triggers behavioural and emotional responses aimed at restoring proximity [18].

The fourth cluster reflects that loss-related memories can be easily triggered by a variety of internal and external cues. This can include specific reminders of the loss, such as time markers, illness-related cues, or internal emotional or physical states [11]. In addition, positive or affiliative cues, such as a shared piece of music, a familiar place, or seeing other families together can also serve as triggers because they activate loss memories through their close association with important experiences in connection to the loss [5].

Fifth, a predominance of negative memories and affect was evident in interview participants’ descriptions of loss-related memories, with even positive memories giving rise to painful emotions [11]. This maps onto the retrieval bias for memories involving the deceased [12], and findings that positive memories of the deceased are also associated with painful reactions [8, 10].

Sixth, some individuals described difficulty retrieving memories that did not include the deceased and reported feeling disconnected from their past self [11]. This disconnection from memories of past self resembles the idea that when the self is strongly defined in relation to the deceased, memory recall is biased towards memories that include the deceased [4].

Some of these memory characteristics appear to be related to, and raise possible hypotheses on mechanisms, for some of the symptoms of PGD. For example, identity disruption, the feeling of having lost a part of one’s self in the diagnostic criteria of PGD, may in part be related to the OG-M cluster disconnection from memories of the past self. Empirical evidence supports the distinction between symptoms and memory characteristics. In cross-lagged models across three time points up to 18 months post-loss, memory characteristics predicted subsequent PGD symptom severity after controlling for prior levels of both symptoms and memory characteristics [16]. If the OG-M were simply capturing PGD symptoms, this residual predictive power would not be expected. In longitudinal mediation models, the effect of memory characteristics on PGD symptoms was mediated by unhelpful coping strategies such as rumination and avoidance [19], a pattern consistent with memory characteristics as upstream maintaining mechanisms rather than symptoms themselves. More recently, memory characteristics were found to be the most prominent variable in a network analysis of cognitive behavioural mechanisms of prolonged grief [20]. Together, these findings suggest that loss-related memory characteristics are closely linked to, but not reducible to, PGD symptoms.

Open questions still remain regarding the causal relationships among individual memory characteristics and PGS. Identifying which aspects of loss memories drive the development of PGS can inform interventions and the identification of individuals who are at risk for PGD. The network approach to psychopathology [21] is uniquely suited to unravel such relationships because it conceptualizes mental disorders as arising from direct causal connections among symptoms and related phenomena.

The Present Study

Several network analyses have been conducted on prolonged grief and related symptoms [22]; however, most of these investigations were cross-sectional. While central symptoms (i.e., symptoms with many and/or strong connections to other symptoms) in networks are candidate intervention targets [23], it is important to first identify the directionality of the relationship (i.e., a highly central symptom can either influence many other symptoms or be influenced by many other symptoms [24]). Dynamic network analyses address this limitation by modelling lagged, directional associations between variables. Thereby, they can reveal which variables drive change in others over time.

To date, only very few temporal network analyses have examined the dynamics among PGS [25, 26]. More broadly, dynamic network analysis can be employed to investigate how proposed mechanisms relate to individual symptoms across time [24, 27, 28], and to what extent, symptoms can influence themselves over time [24, 28]. Dynamic modelling approaches such as the panel graphical vector autoregressive (GVAR) model [29, 30], further allow to disentangle within-person and between-person effects. Whereas between-person effects describe differences across individuals, within-person effects capture fluctuations relative to a person’s own average and are particularly informative for identifying mechanisms and potential treatment targets [24, 31].

The present study therefore aimed to explore the dynamic associations between loss-related memory characteristics and PGS. We were particularly interested in within-person associations across the early bereavement period. Specifically, we investigated the role of the six memory characteristics introduced above – re-experiencing qualities, visceral consequences, proximity seeking, triggers, predominance of negative memories and affect, and disconnection from memories of past self – in relation to PGS assessed during the first 12–18 months after bereavement.

Methods

Participants and Procedure

This study involved 275 bereaved adults recruited within the first 6 months following the death of a close person and reassessed at 6 and 12 months later (T1 = 0–6 months post-loss; T2 = 6–12 months; T3 = 12–18 months). Sociodemographic and loss-related characteristics are provided in Table 1. Participants were recruited online in the UK through bereavement charity email lists, social media advertisements, and keyword-based advertising on the Google content network; all assessments were administered via online surveys and followed ethical guidelines [32]. The survey was available in English and data were collected between 2015 and 2017. The design used fixed 6-month lags between waves (consistent with best practices in longitudinal network modelling) making it compatible with panel network approaches that assume equidistant observation intervals to interpret lagged effects.

Table 1.

Sample characteristics (N = 275)

​	n (%)	M (SD), range
Demographic characteristics
Age, years	​	46.43 (13.2)
Gender
Female	217 (79)	​
Male	58 (21)	​
Education
No qualifications	9 (3.3)	​
High school education	88 (32.0)	​
University degree	110 (40.0)	​
Postgraduate degree	68 (24.7)	​
Employment Status
Student	14 (5.1)	​
Employed full-time	113 (41.1)	​
Employed part-time	37 (13.5)	​
Not currently employed	34 (12.4)	​
On sick leave	13 (4.7)	​
Other	64 (23.3)	​
Estimate household income
Less than 10,000/year	30 (10.9)	​
10,000–24,999	74 (26.9)	​
25,000–49,999	84 (30,5)	​
50,000–74,999	30 (10.9)	​
75,000–99,999	16 (5.8)	​
Over 100,000	16 (5.8)	​
Other	25 (9.1)	​
Loss-related characteristics
Time since loss at T1, months	​	2.94 (2.0), 0–8
Cause of death	​	​
Violent	25 (9)	​
Non-violent	250 (91)	​
The deceased was
Partner	82 (30)	​
Child	25 (9)	​
Parent	104 (38)	​
Sibling	17 (6)	​
Other relative	39 (14)	​
Close non-relative	8 (3)	​
Probable PGD at T3a
No	220 (80.0)	​
Yes	21 (7.6)	​

T1 = 0–6 months post-loss. T3 = 12–18 months.

^a12.4% (n = 34) missing values.

Measures

Prolonged Grief Symptoms

PGS were assessed using an adapted version of the PG-13 [33] that maps onto the DSM-5-TR conceptualisation of PGD. For the present analyses, we focused on 10 symptoms (i.e., yearning, preoccupation with thoughts or memories of the deceased, intense emotional pain, avoidance of reminders, role confusion, difficulty accepting the loss, difficulty moving on, emotional numbness/disbelief, life feeling empty/meaningless, and detachment from others). Items were rated on a Likert scale (1–5), with higher scores reflecting greater symptom severity. Internal consistency in the sample was excellent (α = 0.90).

Memory Characteristics

Loss-related memory properties and more general memory processes were measured with the OG-M [16]. Participants were instructed to respond with reference to their memories of the loss. All items were rated on a scale from 0 = not at all to 4 = very strongly and pertain to the six theory-informed clusters:

• 1.Re-experiencing qualities (7 items) measures frequency and distress of intrusive memories, degree of reliving, nowness, vividness, emotional re-experiencing, and feelings of unreality (e.g., “To what extent did it seem to be happening now instead of being something from the past?”) (α = 0.82).
• 2.Visceral consequences (3 items) assessing bodily pain and fatigue associated with loss memories (e.g., “The memories of [–]’s death make my body ache with overwhelming fatigue”) (α = 0.86).
• 3.Proximity seeking (2 items) measuring urges to comfort or search for the deceased triggered by memory activation (e.g., “I find myself suddenly overcome to find [–]”) (α = 0.74).
• 4.Triggers (6 items) measuring breadth and frequency of cues eliciting memories, including everyday stimuli and calendar reminders (e.g., “Many different things trigger distressing memories of the loss”) (α = 0.86).
• 5.Predominance of negative memories and affect (6 items) which describes bias toward painful recall and difficulty accessing positive affect in positive memories (e.g., “When I try to remember good times we shared, memories of the loss pop up”) (α = 0.84).
• 6.Disconnection from memories of past self (3 items) assesses diminished access to self-related memories independent of the deceased (e.g., “I hardly remember anything that I did without [–]”) (α = 0.80).

Statistical Analysis

To examine the potential conceptual overlap between loss-related memory characteristics and PGS, we conducted confirmatory factor analyses at the item level. Specifically, we compared a one-factor model, in which all memory and symptom items loaded onto a single latent construct, with a two-factor model in which memory characteristics and PGS were specified as related but distinct latent factors. These analyses were conducted as a robustness check to evaluate whether the constructs assessed in the main analyses could be empirically distinguished, given their conceptual proximity. Models were estimated using robust maximum likelihood (MLR), which is appropriate for mixed-scale indicators and provides standard errors and fit indices robust to non-normality [34]. Model fit was evaluated using multiple indices, including the comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation, and standardised root mean square residual, as well as information criteria (AIC, BIC) [35–37]. Differences in model fit were assessed using changes in fit indices and log-likelihood values.

In network analysis, statistical relationships (represented by edges) between variables (represented by nodes) are estimated [38]. Data were analysed using a panel GVAR model to decompose (i) temporal (lagged) relations across waves, (ii) contemporaneous associations within waves, and (iii) between-person effects (modelled to partial out stable inter-individual differences so that temporal and contemporaneous parameters reflect average within-person dynamics). The analyses scripts were adapted from code provided in Freichel et al. [27]. We modelled the network structure of the six theory-informed memory characteristics clusters and the ten PGS. PGS were entered as raw Likert scores and memory characteristics cluster scores were computed as item sums with higher values indicating stronger endorsement of the characteristic.

In this framework, the temporal component is a vector autoregression model comprising autoregressive and cross-lagged paths, whereas the within-wave residual structure is modelled as a Gaussian graphical model of conditional (partial) associations [39]. This specification is related to a stationary random-intercept cross-lagged formulations designed to separate within- from between-person effects, with within-person effects being frequently highlighted as the clinically relevant target of inference [24, 31, 39–41]. The between-person structure is retained for identification, enabling the isolation of trait-like between-person from state-like within-person processes, the latter of which this mechanistic study is interested in [31, 39]. In the present context, the temporal network estimates how deviations from within-person means (across the sample; termed average within-person effects) on a given symptom or memory characteristic at a given time point predict change in another node at the next time point (i.e., 6 months later; e.g., from T1→T2 to T2→T3), adjusting for the target node’s prior level and all other predictors in the system. The contemporaneous network captures how variables co-vary within the same wave after conditioning on all other nodes. The between-person component reflects stable individual differences (e.g., whether people who typically score higher on one construct also tend to score higher on another compared to their peers); it is modelled to help isolate the more clinically pertinent within-person dynamics and is thus not the focus of interpretation in this study [28].

Prior to modelling, we detrended variables for linear time trends and standardized scores across waves to support the assumption of stationarity and to place nodes on comparable scales for network estimation [24, 42, 43]. Missing data were addressed using full information maximum likelihood, which provides unbiased parameter estimates under missing-at-random assumptions and is a state-of-the-art approach to approaching missing data in longitudinal network applications [30, 44].

We first estimated a saturated lag-1 panel GVAR including all possible edges, then applied pruning and step-up model search (two-sided α = 0.05) to obtain sparse, interpretable networks that limit false positives while preserving fit, following current recommendations for network model selection [45]. Model adequacy was evaluated with conventional indices: root mean square error of approximation <0.90, CFI >0.90, TLI >0.90 [36], validated for network modelling [46]. We summarized the temporal network’s dynamics with in-strength (i.e., the degree to which a node is predicted by others at the previous wave) and out-strength (i.e., the degree to which a node predicts others at the next wave) plots, excluding autoregressive effects [38]. The between-subject network was modelled using a Cholesky decomposition which ensures that the eigenvalues of the estimated variance covariance structure are non-negative, leading to credible contemporaneous and temporal networks [47]. All analyses were conducted in R (version 2025.09.0+387) using the psychonetrics package for estimation and qgraph for visualization [48–50].

Results

Confirmatory Factor Analysis

Confirmatory factor analyses indicated that a two-factor model of memory characteristics and PGS provided a better fit than a one-factor model, although improvements were modest at time point 1 (ΔCFI = 0.012) and larger at the final time point (ΔCFI = 0.030). Latent correlations were high (r = 0.93 T1; r = 0.88 at T3), indicating substantial shared variance but enhanced differentiation as grief moves from the acute to the prolonged grief phase. Full confirmatory factor analysis results are reported in the online supplementary materials (for all online suppl. material, see https://doi.org/10.1159/000553394).

The Temporal Network

The saturated model showed an excellent fit to the data (RSMEA = 0.046, CFI = 0.94, TLI = 0.90). The pruned model also revealed an excellent fit (RSMEA = 0.040, CFI = 0.94, TLI = 0.93) and a chi-square difference test indicated that the pruned model was not a significantly worse fit than the saturated model χ²(327) = 366.13, p = 0.07; therefore, the pruned model was adopted in the interest of parsimony.

In the temporal network (Fig. 1), several cross-lagged associations emerged between loss memory characteristics and PGS. Most importantly, re-experiencing qualities (OG-M) had three direct edges, making it the most temporally influential node. Re-experiencing qualities (OG-M) and emotional numbness (PGS) reciprocally predicted increases in each other over time, representing a vicious cycle. Re-experiencing qualities (OG-M) also predicted subsequent difficulty accepting the loss (PGS) and increases in memory triggers (OG-M). In turn, memory triggers (OG-M) predicted elevated difficulty moving forward (PGS). Furthermore, a predominance of negative memories and affect (OG-M) was associated with a pattern of other problems. First, it directly predicted increased yearning (PGS) at the next time point. Second, it also predicted later disconnection from memories of past self (OG-M) which, in turn, predicted increased proximity seeking (OG-M), which subsequently predicted confusion about one’s role in life (PGS). While these associations form a sequential pattern across time, they reflect distinct pairwise predictive effects. Finally, memories with strong visceral consequences (OG-M) predicted increased feelings of loneliness and detachment (PGS). Overall, this pattern of results indicated that memory characteristic nodes predicted PGS at a later time point and not vice versa.

Fig. 1.

Temporal network showing the directed associations of loss memory characteristics and PGS across time. Green circles (nodes) represent memory characteristics and orange nodes represent PGS. Blue lines indicate positive and red lines negative associations (edges), with thickness of the edges indicating the strength of the association. Arrows indicate effects from one node to another at the next time point. Only significant associations between the nodes are shown (p < 0.001), while controlling for all other nodes in the network.

The temporal network revealed few associations within PGS. Loneliness/detachment predicted reduced difficulty accepting the loss, and preoccupation with the deceased predicted increases in later feelings of meaninglessness. Emotional numbness predicted itself over time, thus showing an autoregressive effect. This effect was negative such that higher scores at one time point predicted lower scores at the following time point. Two symptom nodes, avoidance and emotional pain, were not predicted by any other nodes in the network and appeared on the periphery.

The in- and out-strength centrality estimates are shown in Figure 2 for the temporal network. In line with the pattern revealed by the edges, the most influential (out-strength) nodes in the network were memory characteristics, particularly re-experiencing qualities. The nodes most strongly predicted by other nodes (in-strength) were PGS, especially difficulty accepting the death, yearning, difficulty moving forward, and feeling alone or detached from others, and were predominantly predicted by memory characteristics at the prior time point.

Fig. 2.

Out- and in-strength centrality of nodes in the temporal network. Out-strength represents the sum of all outgoing absolute edge weights from a node; in-strength represents the sum of all incoming absolute edge weights to a node. PGS: YEARN, yearning; PREOC, preoccupation; PAIN, emotional pain; AVOID, avoidance; ROLE, role confusion; ACCEPT, difficulty accepting; MOVE, difficulty moving forward; NUMB, emotionally numb; MEANL, meaningless; ALONE, feeling alone or detached. Memory characteristics: RQUAL, re-experiencing qualities; VC, visceral consequences; TRIG, triggers; PNA, negative predominance; DISC, disconnection from memories of past self; PRSEEK, proximity seeking.

The Contemporaneous Network

Associations between nodes in the network within the same time point controlling for all other nodes are shown in Figure 3. All nodes showed positive relationships indicating that they enhance each other within the same time point. The memory characteristics nodes were strongly interconnected: the strongest associations were between re-experiencing qualities and triggers and between triggers and visceral consequences. The strongest edges among PGS variables were observed between yearning, preoccupation with thoughts or memories of the deceased, and emotional pain. The most connected node in the network was a memory characteristics node: re-experiencing qualities, linking to seven nodes in total. These were three other memory characteristics nodes, i.e., predominance of negative memories and affect, triggers, and disconnection from memories of past self and four PGS nodes, i.e., preoccupation, difficulty accepting the death, emotional numbness, and avoidance of reminders of the loss. The next most connected node was the PGS meaninglessness of life with six associations: five being other symptoms variables (role confusion, difficulty accepting the death, difficulty moving forward, emotional numbness, feeling alone or detached) and one memory characteristic (disconnection from memories of past self).

Fig. 3.

Contemporaneous network of associations among memory characteristics and PGS within the same time point. Blue lines indicate positive associations and line thickness represents the strength of the association. Only significant associations between the nodes are shown (p < 0.001), while controlling for all other nodes in the network and controlling for temporal effects.

Discussion

This study investigated the dynamic associations of theory-informed loss memory characteristics and PGS in the first 12–18 months after bereavement. Using a dynamic network model (panel GVAR), we extracted contemporaneous and temporal networks that examined the associations of these variables within the same time point and over time.

Effects of Memory Characteristics over Time

The most striking feature of the temporal network was that memory characteristics predominantly predicted PGS at the subsequent time points. Only the symptom emotional numbness predicted a memory characteristic, re-experiencing qualities, in a vicious cycle: uncontextualised, sensory-rich loss memories appear to promote later emotional numbing, while emotional numbing in turn predicts the persistence of such intrusive memory qualities. Emotional numbness also emerged as an influential symptom in a temporal network analysis of daily dynamics among PGS and has been linked to emotional suppression as a form of avoidance [25]. Our finding thus converges with theoretical models of PGD and PTSD, which propose that avoidance and emotional suppression inhibit memory contextualisation, maintaining distress and dysregulation [5–7]. Re-experiencing qualities were the most influential memory characteristic and also predicted subsequent difficulty accepting the loss, supporting that uncontextualised memories may hinder belief updating, reinforce states of disbelief, and hinder adaptation to a life without the deceased [6, 7].

Predominance of negative memories and affect predicted increased yearning, the only temporal association involving this symptom. This is notable given that yearning is one of the two core symptoms of PGD, and the finding may shed light on its underlying mechanism. Yearning has been conceptualized as including imagery of being reunited with the deceased, accompanied by a bittersweet affective experience [18, 51]. Future research needs to investigate whether negatively biased recall may be a driver of this state, with envisioning reunion in an alternative present or future serving as a psychological counterweight to painful memories. Negative predominance of memories and affect also predicted disconnection from memories of past self, which in turn predicted proximity seeking, pointing to a pattern in which a painful memory bias undermines self-continuity and promotes urges to restore closeness to the deceased. Proximity seeking predicted subsequent feelings of confusion about one’s role in life or that a part of oneself has died. Previous research has demonstrated a relationship between proximity seeking and higher PGS severity [52, 53]. The present study extends these findings by showing that search and rescue urges that inevitably result in a confrontation with the permanent separation may serve to highlight the disruption stemming from the deceased’s absence. This experience can undermine an individual’s sense of who they are, especially in cases where the deceased is central to self-identity. This is in line with the theoretical assumption that proximity seeking reinforces an identity that is entwined with the deceased and inhibits re-organization [4].

Finally, visceral consequences of memories and triggers emerged as important maintaining factors of PGS. Memory triggers predicted increased difficulty moving forward, indicating that memories that are easily triggered by internal and external cues may obstruct planning and re-engagement with future-oriented activities. This is meaningful as fearing that you will be overwhelmed with triggers in a particular situation is likely to raise anxiety and result in avoidance behaviour which thus obstructs processing of the loss memory [5, 6]. Similarly, visceral consequences of loss memories predicted heightened loneliness and detachment, suggesting that memories that include embodied grief reactions are associated with exacerbating perceptions of isolation and separation from others. This may be because grievers worry that the nature of their grief sets them apart from others or that others will not understand or judge them for these reactions. This is in line with evidence suggesting that negative interpretation of others’ reactions to grief expression is a predictor of social disconnection in PGD [54]. Avoidance of activities and social withdrawal is also assumed to be related to poor integration and elaboration of the loss memory and it contributes to the development and maintenance of PGD symptoms [6, 55, 56]. The strong predictive effects of re-experiencing qualities and triggers also map onto work in PTSD, showing that these memory qualities predict PTSD in prospective studies of trauma survivors [57] and that changes in these qualities drive subsequent symptom change in cognitive therapy for PTSD [58, 59].

Together, these findings lend empirical support to theoretical models that assign a central role to memory processes in the development and maintenance of PGD. The result that specific memory features are associated with sustaining and reinforcing core difficulties in adaptation following bereavement is consistent with the notion that how the loss is encoded shapes the trajectory of grief reactions [4–6].

Effects of PGS over Time

The temporal network indicated limited autoregressive effects among PGS and memory characteristics. Only one variable predicted itself across time, suggesting relatively low temporal stability of individual symptoms within this bereaved sample. This may reflect the natural course of symptom fluctuation and partial resolution in the first year following loss, consistent with longitudinal work indicating that many bereaved individuals experience gradual symptom improvement or remain relatively resilient in the face of loss [60, 61]. Interestingly, emotional numbness was the only symptom showing significant autoregressive effects, but in a negative direction, whereby higher levels of numbness at one time point predicted lower levels at the subsequent assessment. This pattern may point to a recovery-related process, with emotional blunting serving as a transient coping response that diminishes over time. An alternative explanation for a negative autoregressive effect may be a state-like feature that varies according to contextual demands [24]. For example, numbness may be heightened during anniversaries, reminders, or stressful life events but resolves when circumstances shift. In this sense, the negative autoregression could indicate that high levels of numbness at one time point are less likely to persist simply because the experience tends to vary naturally over time and context, rather than being a steadily maintained symptom. However, the vicious cycle between emotional numbness and re-experiencing qualities of memories seems to negate that possibility, evidencing a more mechanistic cycle rather than a state that comes and goes in response to context.

Several other temporal pathways highlight the ways in which specific grief symptoms interact dynamically. Loneliness and detachment from others predicted reduced difficulty accepting the loss over time. One possible interpretation is that experiences of social disconnection paradoxically reinforce awareness of the irreversibility of the loss: when the bereaved person feels detached from social life, the absence of the deceased may become more salient, confronting them with the permanence of the loss in ways that facilitate its processing and cognitive acceptance. Particularly in the relatively early bereavement period captured here, detachment from others may provide space to acknowledge and feel the pain of the loss without distraction, for example, through social engagement, consistent with the dual process model’s emphasis on oscillation between loss- and restoration-oriented coping [62]. In this sense, loneliness at one time point may reflect inward emotional processing that, when not extended, supports rather than impedes acceptance over time. Alternatively, loneliness may prompt behavioural adaptation, such as seeking new social connections, which could indirectly support acceptance processes. Another notable pathway was from preoccupation with the deceased to increased feelings of life being meaningless. This suggests that intense cognitive focus on the deceased and, by extension, on the past diverts attention away from present and future goals and contributes to existential emptiness.

Associations within the Same Time Point

The contemporaneous network revealed strong positive concurrent associations among PGS and memory characteristics within the same time point. The core symptoms of yearning and preoccupation, together with emotional pain, formed a tightly interrelated cluster representing separation distress, similar to associations reported by Pociūnaitė-Ott et al. [25]. Of interest, visceral consequences showed widespread associations with PGS, including emotional pain, underscoring that the pain of the loss can be felt physically and consistent with evidence that somatic symptom distress is frequently experienced by patients with PGD [63]. It also supports the proposition that original bodily responses become encoded within the associative memory network and get reactivated as part of the memory [5].

Avoidance of reminders was relatively isolated in the networks, with no connections in the temporal network and only few connections in the contemporaneous network. Similar patterns have emerged in other network analyses of PGD symptoms [64, 65].

Clinical Implications

The present results support loss memory characteristics as a mechanism behind PGS. Given that temporally influential nodes are considered promising candidates for intervention [23], the predominant role of memory characteristics in the temporal network supports the hypothesis that targeting memory processes during treatment will be associated with reducing prolonged grief symptomatology. During grief-focused cognitive behavioural therapy, which is successful in treating prolonged grief [66], exposure exercises typically focus on achieving an integration of the loss into the autobiographical memory and reducing maladaptive avoidance [6, 66]. They target loss-related memories through telling the story of the loss, written accounts of memories, or (imaginal) exposure to important moments connected to the loss [5, 67–70]. Future research needs to unravel which changes in memory characteristics these interventions cause and how they may relate to symptom change over time.

More specifically, the current findings can be linked to the mechanisms targeted in Internet-delivered cognitive therapy for prolonged grief disorder (iCT-PG), which has shown encouraging early results [71]. This treatment addresses cognitive mechanisms and explicitly the memory characteristics featured in the present analysis. For example, a key feature of the memory work done in iCT-PG is a focus on telling the story of the relationship prior to any narrative account of the loss memory. This process, which maps onto the predominance of negative memories and affect, encourages patients to deliberately and effortfully retrieve and elaborate positive memories of the relationship and serves to make these positive representations more accessible once work with the loss memory begins. When the loss memory itself is addressed, patients are guided to produce a detailed narrative account of the death and its surrounding circumstances, with the aim of identifying the worst meanings linked to the most distressing moments. The therapist then supports the patient in generating and integrating information that contradicts or challenges these worst meanings [71], including an emphasis on the enduring positive qualities of the deceased, and the ways in which these qualities remain present in, or can be carried forward by, the patient. This process aims to promote a sense of ongoing connection and continuity of the deceased in the present moment and thus directly targets the need to restore proximity. From a mechanistic perspective, this therapeutic sequence reflects the process of memory updating derived from procedures used in CT-PTSD [72] that helps link the more helpful meanings with the worst moment at an emotional level and reduces re-experiencing.

Another intervention that aims to modify the meaning of aversive memories is imagery rescripting [73–76]. ImRs involves guiding patients to relive the beginning of a memory until the worst moment is about to occur, at which point the sequence of events is altered to produce a more benign outcome that fulfils previously unmet needs. Evidence suggests that ImRs can change the meaning of a memory and enhance a sense of mastery [77, 78], and it has recently been proposed as a suitable treatment for PGD [79]. Further research is needed to establish how the specific memory characteristics highlighted in the present results change during treatments and influence their outcomes.

Limitations

Several limitations must be considered when interpreting the results. First, we did not differentiate between losses due to violent versus non-violent causes of death, although this may influence the characteristics of loss-related memories. Second, the DSM-5-TR criteria for PGD needed to be approximated, as data collection took place before the criteria were finalized. Although our operationalization captured the core features of the disorder, future research should replicate the findings. Third, our sample reflects community bereavement rather than a clinical sample, which limits the generalizability of the findings to PGD. DSM-5-TR diagnostic criteria could only be fully applied at T3 with 7.6% of participants meeting criteria for probable PGD. This percentage is slightly higher than reported in representative studies [1]. Replication in clinical samples is needed to establish whether the processes identified here operate with equal strength in individuals meeting diagnostic criteria for PGD, and the present findings provide a theoretically grounded basis for such work. At the same time, the community sample also constitutes a strength in the present context, as it allows investigation of early mechanistic processes that may differentiate adaptive from maladaptive grief trajectories across the bereavement period. Fourth, all variables were assessed using self-report measures, which precludes to determine whether participants fully met criteria for PGD. Several of our constructs were also modelled as sum scores, unable to account for measurement error. Future applications combining latent variable and network approaches to address measurement error are warranted [80]. Lastly, the sample was predominantly female (79%). Female gender is associated with higher PGD symptom severity [81] which potentially drew more women to the study. However, overrepresentation of women is a structural challenge consistent with the broader grief research literature [56]. Our findings may be most representative of female grief experiences: given the higher risk for PGD and a mounting rather than decreasing symptom trajectory over early bereavement [82], the strength and longitudinal shape of the identified associations may differ in male or more gender-balanced samples. Future research should therefore not only replicate these findings in more balanced samples but also examine sex and gender as moderators of pathways from memory characteristics to symptoms which could directly inform whether memory-focused interventions require tailoring across gender.

Future research would also benefit from incorporating comorbid symptom nodes into the network. Given that PTSD frequently co-occurs with PGD [83] and that memory characteristics play a central role in both disorders, examining how memory characteristics differentially predict PGD and PTSD symptoms and which processes may drive their comorbidity represents a promising avenue for further investigation. Including PTSD symptoms as additional nodes in the present networks was not feasible given sample size constraints.

Conclusion

The present study examined temporal relations between loss-related memory characteristics and PGS across the first 18 months of bereavement using dynamic network analysis. Memory characteristics predominantly impacted PGS over time rather than vice versa, suggesting their role as potential emergent and maintaining mechanisms of PGD symptomatology and supporting a focus on loss-related memories during PGD treatment. These results support theoretical models that propose memory characteristics as one important mechanism in PGD, extend findings regarding re-experiencing qualities from PTSD to PGD, and highlight the need to understand the role of memory work during PGD treatment.

Acknowledgments

Use of generative AI: during the preparation of this work, the authors used Claude (Claude Sonnet 4.6, Anthropic, 2025) in order to improve clarity and grammar of selected phrases. All AI-assisted edits have been reviewed, and the authors take full responsibility for the content.

Statement of Ethics

This study was reviewed and approved by the University of Oxford Medical Sciences Inter-Divisional Research Ethics Committee, Approval No. MS-IDREC-C1-2015-230; MS-IDREC-C1-2015-231. It was carried out in accordance with the Declaration of Helsinki (World Medical Association, 2013). Written informed consent was obtained from all participants prior to their participation in the study.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This work was supported by the Medical Research Council (MR/V001841/1) and the Wellcome Trust (AE:200796); the National Institute for Health Research (NIHR) Biomedical Research Centre, based at Oxford University Hospitals NHS Trust (K.V. Smith), and the NIHR Oxford Health Biomedical Research Centre (K.V. Smith and A. Ehlers). The views expressed are those of the authors and do not necessarily reflect those of the NIHR or the Department of Health and Social Care. The funding sources played no role in the study design, execution and analysis, as well as manuscript conception, planning, writing, and decision to publish.

Author Contributions

Conceptualization: K.V.S. and F.L.-M.; methodology: K.V.S., F.L.-M., and O.V.E.; data curation, investigation, and formal analysis: K.V.S.; resources and funding acquisition: K.V.S. and A.E.; writing – original draft preparation: K.V.S. and F.L.-M. Manuscript revision and editing: K.V.S., F.L.-M., O.V.E., and A.E. All authors have read and agreed to final version of the manuscript.

Funding Statement

This work was supported by the Medical Research Council (MR/V001841/1) and the Wellcome Trust (AE:200796); the National Institute for Health Research (NIHR) Biomedical Research Centre, based at Oxford University Hospitals NHS Trust (K.V. Smith), and the NIHR Oxford Health Biomedical Research Centre (K.V. Smith and A. Ehlers). The views expressed are those of the authors and do not necessarily reflect those of the NIHR or the Department of Health and Social Care. The funding sources played no role in the study design, execution and analysis, as well as manuscript conception, planning, writing, and decision to publish.

Data Availability Statement

The ethical approval for this work stipulated that no data could be disseminated outside the research team. The relevant Ethics Board can be contacted at ethics@medsci.ox.ac.uk, citing MS-IDREC-C1-2015-231 for further information. Readers who wish to access the analysis scripts and corresponding outputs may request them from the corresponding author.