Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories

# Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories
Source: [https://arxiv.org/html/2608.19621](https://arxiv.org/html/2608.19621)
Yujia Zhou\\correspondingBangde DuWeihang SuXinyuan CaoQingyi PanQingyao Ai\\correspondingYueyue WuMin ZhangYiqun Liu

###### Abstract

Large language models \(LLMs\) offer a scalable approach to social simulation, but their credibility depends on how agents are constructed\. Existing methods can partially reproduce population\-level patterns, yet often fail to capture human\-like diversity\. Our analysis shows that static\-profile agents exhibit stronger demographic separation and within\-group compression than humans, a pattern consistent with identity essentialism:demographic labels can encourage models to treat group\-average tendencies as individual traits, homogenizing responses within groups\. We argue that this limitation arises from two related factors: sparse, static agent representations and the limited ability of prompt\-only memory to persistently integrate experience\. Inspired by complementary memory systems, we propose LifeMem, a longitudinal memory framework that combines structured life\-event retrieval with agent\-specific parametric memory for experience integration\. Experiments on Add Health and Understanding Society with three LLMs show that LifeMem improves alignment with human data in terms of response distributions, overall and within\-group diversity, and patterns of within\-person response change across life stages\. These findings highlight the value of longitudinal life\-event memory for constructing more faithful and dynamically evolving social agents\.

Code—https://github\.com/halsayxi/LifeMem

## Introduction

Large language models \(LLMs\) provide a scalable foundation for social simulation, enabling researchers to study human opinions, attitudes, behaviors, and interactions when conventional surveys or experiments are costly or difficult\([2](https://arxiv.org/html/2608.19621#bib.bib1);[3](https://arxiv.org/html/2608.19621#bib.bib2);[35](https://arxiv.org/html/2608.19621#bib.bib23)\)\. The credibility of such simulations depends critically on how agents are constructed\.

Most approaches instantiate agents from demographic labels, personas, or static profiles\([3](https://arxiv.org/html/2608.19621#bib.bib2);[40](https://arxiv.org/html/2608.19621#bib.bib3);[6](https://arxiv.org/html/2608.19621#bib.bib4);[22](https://arxiv.org/html/2608.19621#bib.bib5)\)\. Although these representations effectively condition models on individual attributes, they often produce responses that are less diverse than human data\([6](https://arxiv.org/html/2608.19621#bib.bib4);[44](https://arxiv.org/html/2608.19621#bib.bib18);[48](https://arxiv.org/html/2608.19621#bib.bib16);[45](https://arxiv.org/html/2608.19621#bib.bib42)\)\. This diversity collapse may reflect a form ofidentity essentialism\([44](https://arxiv.org/html/2608.19621#bib.bib18)\)\. In social psychology, essentialism treats social categories as reflecting stable underlying essences, making members of the same category appear fundamentally alike\([38](https://arxiv.org/html/2608.19621#bib.bib17);[5](https://arxiv.org/html/2608.19621#bib.bib28)\)\. In social simulation, a corresponding pattern arises when static identity labels become overly predictive of agent responses, compressing within\-group variation while amplifying between\-group differences\. Figure[1](https://arxiv.org/html/2608.19621#Sx1.F1)presents a response\-space analysis consistent with this pattern\. Such distortions may exaggerate demographic differences, reinforce stereotypes, and bias conclusions drawn from simulated populations\.

![Refer to caption](https://arxiv.org/html/2608.19621v1/intro2.png)Figure 1:Static demographic conditioning produces a pattern consistent with identity essentialism\. We divide 2,000 randomly sampled WVS wave 7 respondents\([19](https://arxiv.org/html/2608.19621#bib.bib47)\)into three socioeconomic status groups and compare human responses with those of profile\-conditioned Llama\-8B agents\. Each individual’s responses are concatenated and projected into a two\-dimensional PCA space\. The separate projections show greater within\-group variation and overlap among humans \(silhouette−0\.02\-0\.02\), whereas agents exhibit stronger within\-group compression and between\-group separation \(silhouette0\.190\.19\)\. Silhouette scores are computed in the original response space before PCA projection\.We attribute this limitation to both the content and form of agent memory\. First, demographic attributes provide only a sparse and static account of an individual, omitting the richer, evolving experiences associated with between\-person differences and within\-person change\([14](https://arxiv.org/html/2608.19621#bib.bib13);[11](https://arxiv.org/html/2608.19621#bib.bib12)\)\. Second, prompt\-based conditioning may lead to diversity collapse\([45](https://arxiv.org/html/2608.19621#bib.bib42)\)\. As shown in Figure[3](https://arxiv.org/html/2608.19621#Sx5.F3), explicit memory alone does not ensure that accumulated experience is persistently integrated into the agent\.

Inspired by cognitive science, we proposeLifeMem, a longitudinal memory framework for constructing dynamically evolving LLM social agents\. Human memory is commonly understood as relying on complementary systems: the hippocampal system rapidly preserves individual experiences, whereas the neocortex gradually integrates information across experiences into persistent representations\([32](https://arxiv.org/html/2608.19621#bib.bib43);[24](https://arxiv.org/html/2608.19621#bib.bib50)\)\. LifeMem translates this distinction into two complementary components for representing life experiences: \(1\) structured life\-event memory, which explicitly preserves event content, timing, and supporting evidence for retrieval; and \(2\) agent\-specific LoRA adapters, which encode accumulated experiences as parametric memory persistent across questions and updated over time\. By combining richer longitudinal information with persistent parametric memory, LifeMem targets limitations in both the content and representation of agent memory, aiming to construct agents that more faithfully reflect individual differences and temporal change\.

We evaluate LifeMem on Add Health\([20](https://arxiv.org/html/2608.19621#bib.bib15)\)and Understanding Society\([8](https://arxiv.org/html/2608.19621#bib.bib14)\)using three instruction\-tuned LLMs from different families:Llama\-3\.1\-8B\-Instruct\(hereafter,Llama\-8B\)\([16](https://arxiv.org/html/2608.19621#bib.bib44)\),Ministral\-3\-8B\-Instruct\-2512\([28](https://arxiv.org/html/2608.19621#bib.bib45)\), andQwen3\.5\-9B\([39](https://arxiv.org/html/2608.19621#bib.bib46)\)\. Compared to static conditioning, diversity\-oriented prompting, non\-parametric memory, and control baselines, LifeMem improves alignment with human data in terms of response distributions, overall and within\-group diversity, and within\-person response change across life stages\. Our contributions are:

- •We identify identity essentialism in static\-profile LLM agents, which reduces within\-group diversity and exaggerates between\-group differences\.
- •We propose LifeMem, combining structured life\-event memory with agent\-specific parametric memory to model diverse and evolving individuals\.
- •Across two longitudinal surveys and three LLMs, LifeMem improves alignment with human distributions, diversity, and within\-person changes\.

## Related Work

Current LLM\-based social simulation often relies on persona conditioning and static agent representations\.LLMs are increasingly used as experimental participants and survey respondents\([2](https://arxiv.org/html/2608.19621#bib.bib1);[3](https://arxiv.org/html/2608.19621#bib.bib2);[40](https://arxiv.org/html/2608.19621#bib.bib3);[6](https://arxiv.org/html/2608.19621#bib.bib4);[22](https://arxiv.org/html/2608.19621#bib.bib5);[30](https://arxiv.org/html/2608.19621#bib.bib6)\)\. Existing approaches often construct agents from interviews, interaction histories, or social\-media records\([36](https://arxiv.org/html/2608.19621#bib.bib25);[13](https://arxiv.org/html/2608.19621#bib.bib11);[27](https://arxiv.org/html/2608.19621#bib.bib26);[12](https://arxiv.org/html/2608.19621#bib.bib35);[18](https://arxiv.org/html/2608.19621#bib.bib36)\), supporting partial alignment with populations\. However, persona conditioning may reduce response variance or introduce bias\([6](https://arxiv.org/html/2608.19621#bib.bib4);[7](https://arxiv.org/html/2608.19621#bib.bib27);[44](https://arxiv.org/html/2608.19621#bib.bib18)\)\. Moreover, static profiles can compress individuals into demographic types\([48](https://arxiv.org/html/2608.19621#bib.bib16);[22](https://arxiv.org/html/2608.19621#bib.bib5)\)\. This pattern resembles identity essentialism, which attributes stable properties to social categories\([38](https://arxiv.org/html/2608.19621#bib.bib17);[5](https://arxiv.org/html/2608.19621#bib.bib28)\), although demographic conditioning is not inherently essentialist\([44](https://arxiv.org/html/2608.19621#bib.bib18)\)\. More broadly, LLM social agents are typically fixed rather than updated from temporally ordered observations\. In contrast, longitudinal research distinguishes between\-person differences from within\-person change and uses detailed life\-event histories to better characterize individual trajectories\([11](https://arxiv.org/html/2608.19621#bib.bib12);[14](https://arxiv.org/html/2608.19621#bib.bib13)\)\.

![Refer to caption](https://arxiv.org/html/2608.19621v1/framework.png)Figure 2:Overview of LifeMem\. \(a\) Individual life trajectories contain diverse longitudinal experiences\. \(b\) Structured memory supports explicit retrieval, while agent\-specific LoRA adapters consolidate experiences into parametric memory through recursive updates\. \(c\) The resulting agent state evolves across life stages\. \(d\) Agents are initialized from similar demographic profiles, which may produce homogeneous opinions\. \(e\) LifeMem produces diverse and dynamic opinions\.Prior work improves diversity at the output, prompt, memory, or parameter level\.Output\-level methods modify sampling, decoding\([21](https://arxiv.org/html/2608.19621#bib.bib29);[37](https://arxiv.org/html/2608.19621#bib.bib7);[26](https://arxiv.org/html/2608.19621#bib.bib30);[43](https://arxiv.org/html/2608.19621#bib.bib31);[10](https://arxiv.org/html/2608.19621#bib.bib32);[47](https://arxiv.org/html/2608.19621#bib.bib9);[49](https://arxiv.org/html/2608.19621#bib.bib10)\), or human involvement\([1](https://arxiv.org/html/2608.19621#bib.bib8)\)to increase response diversity, but the resulting variation is often stochastic rather than grounded in individual experience\. Prompt\-level methods introduce multilingual, anti\-stereotype, biography\-based, population\-aligned, or ensemble personas\([46](https://arxiv.org/html/2608.19621#bib.bib19);[41](https://arxiv.org/html/2608.19621#bib.bib20);[30](https://arxiv.org/html/2608.19621#bib.bib6);[15](https://arxiv.org/html/2608.19621#bib.bib33);[23](https://arxiv.org/html/2608.19621#bib.bib34);[4](https://arxiv.org/html/2608.19621#bib.bib39)\), but they generally lack an evolving representation of individuals\. Memory methods store, retrieve, update, or forget experience\([31](https://arxiv.org/html/2608.19621#bib.bib21);[25](https://arxiv.org/html/2608.19621#bib.bib22);[35](https://arxiv.org/html/2608.19621#bib.bib23);[50](https://arxiv.org/html/2608.19621#bib.bib40);[29](https://arxiv.org/html/2608.19621#bib.bib38);[17](https://arxiv.org/html/2608.19621#bib.bib37)\), but retrieval alone may not support persistent integration\. Parameter\-level methods encode knowledge or identity into model parameters or activations\([42](https://arxiv.org/html/2608.19621#bib.bib24);[9](https://arxiv.org/html/2608.19621#bib.bib41);[45](https://arxiv.org/html/2608.19621#bib.bib42)\), but are not typically updated from longitudinal observations\.

## Methodology

This section formulates the longitudinal simulation task and introduces LifeMem’s dual\-memory framework\. We then describe its structured memory for explicit event retrieval and parametric memory for integrating experiences across waves\.

### Problem Formulation

Letℐ=\{1,…,N\}\\mathcal\{I\}=\\\{1,\\ldots,N\\\}denote a population of individuals and𝒲=\{1,…,T\}\\mathcal\{W\}=\\\{1,\\ldots,T\\\}the ordered longitudinal waves\. Each individualiihas a static profilePiP\_\{i\}constructed from background information and a sequence of life events observed at wavett,

ℰi,t=\{ei,t,k\}k=1Ki,t,\\mathcal\{E\}\_\{i,t\}=\\\{e\_\{i,t,k\}\\\}\_\{k=1\}^\{K\_\{i,t\}\},\(1\)whereei,t,ke\_\{i,t,k\}denotes thekk\-th observed life event andKi,t=\|ℰi,t\|K\_\{i,t\}=\|\\mathcal\{E\}\_\{i,t\}\|\. The accumulated trajectory through wavettis

𝒯i,≤t=⋃τ=1tℰi,τ\.\\mathcal\{T\}\_\{i,\\leq t\}=\\bigcup\_\{\\tau=1\}^\{t\}\\mathcal\{E\}\_\{i,\\tau\}\.\(2\)
For an evaluation questionqt,jq\_\{t,j\}, LifeMem predicts

y^i,t,j=Fθ,Ai,t​\(Pi,Ri,t​\(qt,j\),qt,j\),\\hat\{y\}\_\{i,t,j\}=F\_\{\\theta,A\_\{i,t\}\}\\left\(P\_\{i\},R\_\{i,t\}\(q\_\{t,j\}\),q\_\{t,j\}\\right\),\(3\)whereθ\\thetais the frozen base LLM,Ai,tA\_\{i,t\}is an agent\-specific LoRA adapter, andRi,t​\(qt,j\)R\_\{i,t\}\(q\_\{t,j\}\)contains events retrieved from structured memory\. LifeMem combines question\-dependent evidence with a persistent individual state to model between\-person differences and within\-person change\.

### LifeMem Overview

LifeMem is a dual\-memory framework inspired by complementary learning systems theory, which distinguishes rapid encoding of specific experiences in the hippocampal system from gradual integration in the neocortex\([32](https://arxiv.org/html/2608.19621#bib.bib43);[34](https://arxiv.org/html/2608.19621#bib.bib48);[33](https://arxiv.org/html/2608.19621#bib.bib49)\)\. In LifeMem, structured memory stores explicit and traceable life events, while agent\-specific LoRA adapters integrate accumulated experiences into persistent parametric memory, as illustrated in Figure[2](https://arxiv.org/html/2608.19621#Sx2.F2)\.

Both memories receive the same longitudinal event stream but serve complementary roles\. Structured memory provides relevant evidence for the question, whereas parametric memory integrates accumulated experience across questions and waves without repeatedly adding the full trajectory into the prompt\. Temporal decay prioritizes retrieval, and replay preserves earlier information during sequential adapter updates\.

### Hippocampal Structured Memory

Each life eventei,t,ke\_\{i,t,k\}is derived from a survey question–answer pair and converted into a second\-person statementxi,t,kx\_\{i,t,k\}while retaining its original question and response\. A frozen encodergϕg\_\{\\phi\}maps the statement to an embedding

𝐡i,t,k=gϕ​\(xi,t,k\)\.\\mathbf\{h\}\_\{i,t,k\}=g\_\{\\phi\}\(x\_\{i,t,k\}\)\.\(4\)The structured memory stores each event together with its embedding and wave index:

ℳi,tS=\{\(ei,τ,k,𝐡i,τ,k,τ\)∣1≤τ≤t\}\.\\mathcal\{M\}^\{S\}\_\{i,t\}=\\left\\\{\\left\(e\_\{i,\\tau,k\},\\mathbf\{h\}\_\{i,\\tau,k\},\\tau\\right\)\\mid 1\\leq\\tau\\leq t\\right\\\}\.\(5\)New events are appended without removing earlier records, preserving a traceable account of the observed trajectory\.

For a questionqq, retrieval combines semantic relevance with temporal decay:

Scoret⁡\(q,ei,τ,k\)=sim⁡\(gϕ​\(q\),𝐡i,τ,k\)​exp⁡\[−λ⁡\(t−τ\)\],\\operatorname\{Score\}\_\{t\}\(q,e\_\{i,\\tau,k\}\)=\\operatorname\{sim\}\\left\(g\_\{\\phi\}\(q\),\\mathbf\{h\}\_\{i,\\tau,k\}\\right\)\\exp\[\-\\lambda\(t\-\\tau\)\],\(6\)whereλ≥0\\lambda\\geq 0controls the preference for recent events\. This decay modifies retrieval priority rather than deleting older experiences, which remain available when sufficiently relevant\. The top\-KKevents form the retrieved evidence:

Ri,t​\(q\)=Top−Ke∈ℳi,tS⁡Scoret⁡\(q,e\)\.R\_\{i,t\}\(q\)=\\operatorname\{Top\-K\}\_\{e\\in\\mathcal\{M\}^\{S\}\_\{i,t\}\}\\operatorname\{Score\}\_\{t\}\(q,e\)\.\(7\)The selected statements are inserted into the inference prompt, providing question\-dependent evidence without exposing the complete history\.

### Cortical Parametric Memory

LifeMem maintains an independently updated LoRA adapter for each agent while keeping the base LLM frozen\. For a target layerll, the agent\-specific weight is

𝐖i,t\(l\)=𝐖0\(l\)\+αLoRAr​𝐁i,t\(l\)​𝐀i,t\(l\),\\mathbf\{W\}\_\{i,t\}^\{\(l\)\}=\\mathbf\{W\}\_\{0\}^\{\(l\)\}\+\\frac\{\\alpha\_\{\\mathrm\{LoRA\}\}\}\{r\}\\mathbf\{B\}\_\{i,t\}^\{\(l\)\}\\mathbf\{A\}\_\{i,t\}^\{\(l\)\},\(8\)whererrandαLoRA\\alpha\_\{\\mathrm\{LoRA\}\}denote the LoRA rank and scaling coefficient\. We useΘi,tLoRA\\Theta\_\{i,t\}^\{\\mathrm\{LoRA\}\}to denote all agent\-specific LoRA parameters, including\{𝐀i,t\(l\),𝐁i,t\(l\)\}l\\\{\\mathbf\{A\}\_\{i,t\}^\{\(l\)\},\\mathbf\{B\}\_\{i,t\}^\{\(l\)\}\\\}\_\{l\}\. Separate adapters allow individuals with similar demographic profiles to develop different parametric states as their trajectories diverge\.

Each life event yields two training instances: a survey question–answer pair preserving the observed response and an event\-reconstruction instance linking the event to the agent’s first\-person representation\. Together, they encode the survey signal and individualized semantics\.

Adapters are updated sequentially across waves:

Θi,tLoRA,\(0\)=Θi,t−1LoRA,t\>1,\\Theta\_\{i,t\}^\{\\mathrm\{LoRA\},\(0\)\}=\\Theta\_\{i,t\-1\}^\{\\mathrm\{LoRA\}\},\\qquad t\>1,\(9\)with the first wave initialized from a shared adapter state\. At each wave, training combines instances from current events with a small sample of historical instances sampled uniformly at random for replay:

ℒi,t=∑z∈ℬi,tcurℓ⁡\(z,θ,Θi,tLoRA\)\+η​∑z∈ℬi,trepℓ⁡\(z,θ,Θi,tLoRA\),\\mathcal\{L\}\_\{i,t\}=\\sum\_\{z\\in\\mathcal\{B\}\_\{i,t\}^\{\\mathrm\{cur\}\}\}\\ell\(z;\\theta,\\Theta\_\{i,t\}^\{\\mathrm\{LoRA\}\}\)\+\\eta\\sum\_\{z\\in\\mathcal\{B\}\_\{i,t\}^\{\\mathrm\{rep\}\}\}\\ell\(z;\\theta,\\Theta\_\{i,t\}^\{\\mathrm\{LoRA\}\}\),\(10\)whereη\\etacontrols the contribution of replayed experiences\. The base model remains frozen, and only the agent\-specific low\-rank parametersΘi,tLoRA\\Theta\_\{i,t\}^\{\\mathrm\{LoRA\}\}are optimized\. Replay serves as a lightweight mechanism for mitigating catastrophic forgetting rather than a separate memory component\.

## Experimental Setup

### Datasets

We evaluate LifeMem on the National Longitudinal Study of Adolescent to Adult Health \(Add Health\) and the UK Household Longitudinal Study \(Understanding Society\)\. We use six waves from Add Health and fifteen waves from Understanding Society\. To ensure consistent longitudinal tracking, we retain only respondents observed in all selected waves\.

Survey variables are semantically categorized into three types: demographic attributes for initial profiles, life events describing evolving experiences, and evaluation targets\. To prevent leakage, evaluation targets are excluded from both demographic and life\-event variables\. To evaluate within\-person opinion changes, we further identify a shared set of evaluation questions available in every wave of Understanding Society and use them as the longitudinal evaluation set\.

### Baselines

Static conditioning includesDirectandProfile\([3](https://arxiv.org/html/2608.19621#bib.bib2);[40](https://arxiv.org/html/2608.19621#bib.bib3);[6](https://arxiv.org/html/2608.19621#bib.bib4);[22](https://arxiv.org/html/2608.19621#bib.bib5)\); diversity\-oriented prompting includesMultilingual\([46](https://arxiv.org/html/2608.19621#bib.bib19)\)andAnti\-Stereotype\([41](https://arxiv.org/html/2608.19621#bib.bib20)\); and non\-parametric memory includesSimVBG\([13](https://arxiv.org/html/2608.19621#bib.bib11)\),Full History\([31](https://arxiv.org/html/2608.19621#bib.bib21)\), andEvent RAG\([25](https://arxiv.org/html/2608.19621#bib.bib22);[31](https://arxiv.org/html/2608.19621#bib.bib21)\)\.Random Eventuses random events to test the importance of respondent\-specific trajectories\.

### Evaluation Metrics

We use four lower\-is\-better metrics\.*KL divergence*measures alignment between human and agent response distributions\. The*within\-group pairwise distance gap*is the absolute difference between human and agent within\-group pairwise response distances\. A smaller gap indicates that simulated within\-group diversity is closer to that of humans, suggesting less identity\-essentialist behavior\. The*normalized entropy gap*is the absolute difference between the normalized response entropies of humans and agents\. The*transition\-distribution JS divergence*measures how closely the model reproduces the human population\-level distribution of response transitions between adjacent survey waves\.

### Implementation Details

We use three similarly sized LLMs: Llama\-3\.1\-8B\-Instruct, Ministral\-3\-8B\-Instruct\-2512, and Qwen3\.5\-9B\. Using a random seed of 42, we sample 100 respondents per dataset\. All methods share the same samples, questions, backbones, and deterministic decoding settings, with temperature00and sampling disabled\.Event RAGuses the stronger bge\-m3 retriever, whereas LifeMem uses the lightweight all\-MiniLM\-L6\-v2 encoder\. Both methods retrieve the top\-K=5K=5events\. LifeMem uses a temporal decay factor of0\.1050\.105, rank\-88LoRA adapters, replay size44, and replay weight0\.50\.5\. Experiments run on one NVIDIA A800\-SXM4 80 GB GPU\.

## Experiments and Results

CategoryMethodAdd HealthUnderstanding SocietyKL Div\.↓\\downarrowWG Gap↓\\downarrowEnt\. Gap↓\\downarrowKL Div\.↓\\downarrowWG Gap↓\\downarrowEnt\. Gap↓\\downarrowTrans\. JS↓\\downarrowLlama\-3\.1\-8B\-InstructStaticConditioningDirect14\.7551∗0\.5596∗0\.7053∗15\.3734∗0\.5561∗0\.7076∗0\.4580∗Profile8\.6719∗0\.3951∗0\.5112∗7\.1522∗0\.3835∗0\.4972∗0\.3886∗Diversity\-OrientedPromptingMultilingual10\.4654∗0\.3733∗0\.5035∗11\.7367∗0\.3446∗0\.4763∗0\.4570∗Anti\-Stereotype9\.8270∗0\.4163∗0\.5452∗8\.1485∗0\.4066∗0\.5120∗0\.3592∗Non\-ParametricMemorySimVBG6\.8058∗0\.3380∗0\.4345∗6\.2912∗0\.3544∗0\.4730∗0\.3709∗Full History6\.1176∗0\.3079∗0\.3974∗4\.9344∗0\.3198∗0\.4004∗0\.3633∗Event RAG5\.8264∗0\.2964∗0\.3951∗5\.2935∗0\.3163∗0\.3991∗0\.3648∗Control BaselineRandom Event6\.8821∗0\.3524∗0\.4538∗5\.5673∗0\.3429∗0\.4539∗0\.3803∗Proposed MethodLifeMem4\.06350\.23090\.32073\.45290\.28860\.37420\.3331Ministral\-3\-8B\-Instruct\-2512StaticConditioningDirect15\.5952∗0\.5679∗0\.7161∗15\.4925∗0\.5648∗0\.7182∗0\.5076∗Profile8\.8060∗0\.4120∗0\.5263∗6\.8298∗0\.3622∗0\.4622∗0\.3609Diversity\-OrientedPromptingMultilingual12\.7028∗0\.3597∗0\.5129∗13\.1824∗0\.3942∗0\.5350∗0\.4889∗Anti\-Stereotype10\.0458∗0\.4339∗0\.5537∗8\.7660∗0\.3946∗0\.5066∗0\.3757Non\-ParametricMemorySimVBG6\.9946∗0\.3427∗0\.4566∗4\.8323∗0\.3272∗0\.4156∗0\.3610Full History4\.7175∗0\.2893∗0\.3817∗3\.8053∗0\.2927∗0\.3413∗0\.3528Event RAG5\.4368∗0\.2849∗0\.3895∗4\.3342∗0\.2887∗0\.3413∗0\.4693∗Control BaselineRandom Event6\.1787∗0\.3578∗0\.4647∗4\.7275∗0\.3156∗0\.4048∗0\.3924Proposed MethodLifeMem2\.29590\.19280\.27831\.96590\.22770\.26590\.3399Qwen3\.5\-9BStaticConditioningDirect13\.9957∗0\.5482∗0\.6923∗15\.9998∗0\.5559∗0\.7080∗0\.4452∗Profile8\.0174∗0\.4004∗0\.5093∗5\.3214∗0\.3463∗0\.4325∗0\.3471∗Diversity\-OrientedPromptingMultilingual7\.7194∗0\.22110\.339510\.9903∗0\.3438∗0\.4728∗0\.4405∗Anti\-Stereotype7\.7407∗0\.3743∗0\.4770∗6\.0185∗0\.3422∗0\.4254∗0\.3356∗Non\-ParametricMemorySimVBG5\.7090∗0\.3307∗0\.4311∗4\.4864∗0\.3182∗0\.3889∗0\.3298∗Full History4\.52660\.25990\.34323\.7659∗0\.2979∗0\.3441∗0\.3203∗Event RAG4\.37010\.26760\.35743\.5507∗0\.2673∗0\.30700\.3394∗Control BaselineRandom Event6\.0741∗0\.3625∗0\.4584∗3\.7848∗0\.3031∗0\.3707∗0\.3466∗Proposed MethodLifeMem4\.11770\.27230\.36012\.68790\.25190\.29320\.3060

Table 1:Overall results on Add Health and Understanding Society across three models\. KL Div\., WG Gap, Ent\. Gap, and Trans\. JS denote KL divergence, within\-group pairwise distance gap, normalized entropy gap, and transition\-distribution JS divergence\. Lower values are better for all metrics\. Bold marks the best result for each model and metric; LifeMem is shaded light gray\.∗indicates results significantly worse than LifeMem in a two\-sided pairedtt\-test conducted over shared questions \(p<0\.05p<0\.05\)\.This section evaluates LifeMem’s effectiveness, mechanisms, scaling, and efficiency\. We examine explicit\-memory limitations, overall performance, parameter differentiation, component contributions, and computational costs\.

### Limits of Explicit Prompt Memory

MethodAdd HealthUnderstanding SocietyKL Div\.↓\\downarrowWG Gap↓\\downarrowEnt\. Gap↓\\downarrowKL Div\.↓\\downarrowWG Gap↓\\downarrowEnt\. Gap↓\\downarrowTrans\. JS↓\\downarrowLlama\-3\.1\-8B\-InstructLifeMem \(Full\)4\.06350\.23090\.32073\.45290\.28860\.37420\.3331LifeMem w/o Param\. Mem\.5\.3517∗0\.2554∗0\.34365\.0874∗0\.3259∗0\.4150∗0\.3594∗LifeMem w/o Struct\. Mem\.6\.2815∗0\.3132∗0\.4164∗4\.7555∗0\.3379∗0\.4318∗0\.3614∗Ministral\-3\-8B\-Instruct\-2512LifeMem \(Full\)2\.29590\.19280\.27831\.96590\.22770\.26590\.3399LifeMem w/o Param\. Mem\.4\.6504∗0\.2516∗0\.3456∗3\.8454∗0\.2810∗0\.3389∗0\.3838LifeMem w/o Struct\. Mem\.3\.3799∗0\.2627∗0\.3597∗2\.4410∗0\.2483∗0\.3093∗0\.3443Qwen3\.5\-9BLifeMem \(Full\)4\.11770\.27230\.36012\.68790\.25190\.29320\.3060LifeMem w/o Param\. Mem\.4\.42950\.27330\.36043\.4224∗0\.2738∗0\.3193∗0\.3411∗LifeMem w/o Struct\. Mem\.6\.0934∗0\.3651∗0\.4760∗3\.2797∗0\.2842∗0\.3430∗0\.3235∗

Table 2:Ablation study of LifeMem on Add Health and Understanding Society\. KL Div\., WG Gap, Ent\. Gap, and Trans\. JS denote KL divergence, within\-group pairwise distance gap, normalized entropy gap, and transition\-distribution JS divergence\. All metrics are lower\-is\-better\. Bold marks the best result for each model and metric; LifeMem \(Full\) is shaded light gray\.∗indicates results significantly worse than LifeMem \(Full\) in a two\-sided pairedtt\-test conducted over shared questions \(p<0\.05p<0\.05\)\.Figure 3:Effect ofEvent RAGretrieval depth on Add Health with Llama\-8B across 100 agents\. We report alignment metrics, runtime, and input tokens for top\-K∈\[1,180\]K\\in\[1,180\]\. Stars mark LifeMem atK=5K=5\. The shaded bands indicate standard deviation across waves\.![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_us_llama.png)Figure 4:PCA visualization of LifeMem’s agent\-specific LoRA states for 100 Understanding Society agents using Llama\-8B\. Points represent states after each wave, with trajectories connecting the same agent over time\. Agent states increasingly diverge as life histories accumulate\.We test whether retrieving more explicit life\-history information is sufficient for longitudinal simulation by varyingEvent RAGretrieval depth fromK=1K=1toK=180K=180\(i\.e\.,Full History\) on Add Health with Llama\-8B\. We evaluate distributional and diversity alignment together with runtime and average input\-token use to characterize both effectiveness and computational cost\.

As shown in Figure[3](https://arxiv.org/html/2608.19621#Sx5.F3), the three simulation metrics initially improve but largely saturate aroundK=40K=40and sometimes deteriorate, while runtime and input\-token use continue to increase\. This trade\-off illustrates a practical limitation of prompt\-only memory and motivates combining selective retrieval with persistent parametric memory\.

### Overall Performance and Diversity

Table[1](https://arxiv.org/html/2608.19621#Sx5.T1)shows that LifeMem performs consistently well across four dimensions, with largely significant gains across three backbones and two datasets\. It reduces distributional and diversity gaps relative to human responses across settings\.Directperforms similarly poorly across models, while Ministral shows the largest improvement with LifeMem\.

Profileimproves overDirectbut remains behind LifeMem, suggesting that static demographic profiles cannot fully capture human heterogeneity and temporal variation in these settings\.MultilingualandAnti\-Stereotypeprompting yield inconsistent gains, suggesting that these prompting strategies do not reliably reproduce human\-like diversity across settings\.SimVBG,Full History, andEvent RAGperform better than static baselines but still generally trail LifeMem, suggesting that non\-parametric memory alone may be insufficient to achieve the same level of alignment\. LifeMem also outperformsRandom Event, highlighting the importance of respondent\-specific trajectories\.

One notable exception is Qwen3\.5\-9B on Add Health, whereMultilingualachieves slightly lower within\-group pairwise distance gap and normalized entropy gap than LifeMem\. However, its substantially higher KL divergence indicates that these gains in diversity alignment are accompanied by poorer agreement with the overall human response distribution, suggesting a less balanced improvement across evaluation dimensions\.

Overall, LifeMem improves alignment with human data across response distributions, overall and within\-group diversity, and longitudinal response dynamics\. These results suggest that LifeMem mitigates patterns consistent with identity essentialism by reducing within\-group homogenization while maintaining closer alignment with human responses\.

### Ablation Study

Table[2](https://arxiv.org/html/2608.19621#Sx5.T2)evaluates the contributions of structured and parametric memory\. Removing either component significantly degrades performance on most metrics, supporting their complementarity\. Parametric memory maintains a persistent agent state that integrates experiences across waves beyond the events retrieved for the current question, whereas structured memory supplies explicit, traceable, and temporally grounded evidence for generation and tends to contribute more strongly to diversity alignment\. Their relative effects vary across different models and datasets\.

### Evolution of Agent\-Specific LoRA States

Figure[4](https://arxiv.org/html/2608.19621#Sx5.F4)uses PCA to project the high\-dimensional agent\-specific LoRA states into a shared two\-dimensional space, illustrating how the states of 100 Llama\-8B agents evolve over 15 waves on Understanding Society\. The initially concentrated adapters become progressively more dispersed and follow distinct trajectories as individual histories accumulate over time\. This pattern suggests that sequential updates gradually encode heterogeneous life experiences into differentiated parametric states, reflecting increasing personalization across agents with different longitudinal trajectories\.

### Scaling with Accumulated Life Experiences

Figure 5:Effect of life\-event coverage in LifeMem on Add Health with Llama\-8B\. We simulate 100 agents while varying the maximum number of life events available per wave from 5 to 180\. Curves report KL divergence, within\-group pairwise distance gap, and normalized entropy gap\. The shaded bands indicate standard deviation across waves\. Increasing life\-event coverage generally improves alignment with human response distributions and diversity\.Figure[5](https://arxiv.org/html/2608.19621#Sx5.F5)examines the effect of longitudinal life\-event coverage in LifeMem on the Add Health dataset with Llama\-8B\. We vary the maximum number of available events per wave from 5 to 180 for 100 agents\. Increasing coverage generally reduces KL divergence, normalized entropy gap, and within\-group pairwise distance gap, with the improvements gradually plateauing after roughly 90 events per wave\. This pattern indicates that richer longitudinal histories provide more informative and individualized evidence for distinguishing individuals and recovering more human\-like response distributions and levels of diversity over time\. Overall, these results further support our motivation to model agents through accumulated life experiences rather than relying on static demographic profiles alone\.

### Efficiency Analysis

All efficiency measurements are obtained under the same hardware configuration\. We focus primarily on relative trends and comparisons across methods, as absolute runtime may vary with different hardware characteristics, including CPU, GPU, memory, and storage configurations\.

Online inference\.Table[3](https://arxiv.org/html/2608.19621#Sx5.T3)compares the average per\-question inference latency across different methods on the Add Health and Understanding Society datasets using Llama\-8B\. Retrieval latency is included for both theEvent RAGand LifeMem methods to ensure a consistent end\-to\-end comparison\. LifeMem is slower than lightweight prompt\-only baselines but remains faster thanFull History,Event RAG, andSimVBGon both datasets\. Adapter loading introduces little additional latency and can be amortized across questions once an agent’s adapter is loaded\.

Table 3:Average per\-question inference latency with Llama\-8B\. AH and USoc denote Add Health and Understanding Society\. Time is reported in milliseconds, and Rel\. denotes latency relative toDirect\. Retrieval latency is included forEvent RAGand LifeMem\. For LifeMem, the subscripted\+x\+xindicates the LoRA adapter\-loading latency\. Rel\. excludes the one\-time LoRA adapter\-loading latency\.Storage\.Figure[6](https://arxiv.org/html/2608.19621#Sx5.F6)compares per\-agent adapter and structured life\-event storage for Llama\-8B as longitudinal histories accumulate\. Structured event storage grows with trajectory length because newly observed experiences must be retained explicitly, whereas each agent maintains a single fixed\-size LoRA adapter that is updated after each wave\. The updated adapter is then used for subsequent inference, so parametric\-memory storage remains bounded rather than growing with the number of accumulated events\.

Figure 6:Per\-agent storage with Llama\-8B as life events accumulate\. The fixed\-size LoRA adapter remains constant, while structured life\-event storage grows with trajectory length\.Offline training\.With Llama\-8B, the average offline adapter update time per agent and wave is 7\.21 seconds on Add Health and 2\.63 seconds on Understanding Society\. This one\-time cost is incurred during wave\-level updates and is not repeated for each subsequent online query\.

## Conclusion

This work aims to mitigate patterns consistent with identity essentialism in LLM social simulation, where static profiles can suppress individual diversity and temporal change\. LifeMem combines structured life\-event retrieval with agent\-specific LoRA parametric memory\. Across two longitudinal datasets and three model backbones, it consistently improves alignment with human response distributions, diversity, and changes across life stages\. Overall, the results suggest that, in these settings, incorporating longitudinal life experiences together with persistent parametric memory provides a more faithful representation of individuals\.

## Ethical Statement

We use only public\-use, de\-identified survey data, comply with applicable data\-use agreements, avoid re\-identification, and report aggregate results\. Because agent\-specific adapters may memorize respondent information, we do not release respondent\-level adapters\. These agents are not digital replicas and should not be used for consequential person\-level decisions; they are intended for methodological research on longitudinal LLM social simulation\.

## Acknowledgements

This work is supported by the Research Project of Quancheng Laboratory, China \(Grant No\. QCL20250105\), the National Natural Science Foundation of China No\. 62502260, and the Postdoctoral Fellowship Program of China Postdoctoral Science Foundation No\. GZC20240833\.

## References

- Abels and Lenaerts \(2025\)A\. Abels and T\. LenaertsWisdom from diversity: bias mitigation through hybrid human\-llm crowds\.arXiv preprint arXiv:2505\.12349\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Aheret al\.\(2023\)G\. V\. Aher, R\. I\. Arriaga, and A\. T\. KalaiUsing large language models to simulate multiple humans and replicate human subject studies\.InInternational conference on machine learning,pp\. 337–371\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Argyleet al\.\(2023\)L\. P\. Argyle, E\. C\. Busby, N\. Fulda, J\. R\. Gubler, C\. Rytting, and D\. WingateOut of one, many: using language models to simulate human samples\.Political Analysis31\(3\),pp\. 337–351\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px1.p1.1),[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px2.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Ashkinazeet al\.\(2025\)J\. Ashkinaze, E\. Fry, N\. Edara, E\. Gilbert, and C\. BudakPlurals: a system for guiding llms via simulated social ensembles\.InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems,pp\. 1–21\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Bastian and Haslam \(2006\)B\. Bastian and N\. HaslamPsychological essentialism and stereotype endorsement\.Journal of experimental social psychology42\(2\),pp\. 228–235\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Bisbeeet al\.\(2024\)J\. Bisbee, J\. D\. Clinton, C\. Dorff, B\. Kenkel, and J\. M\. LarsonSynthetic replacements for human survey data? the perils of large language models\.Political Analysis32\(4\),pp\. 401–416\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px1.p1.1),[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px2.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Boelaertet al\.\(2025\)J\. Boelaert, S\. Coavoux, É\. Ollion, I\. Petev, and P\. PrägMachine bias\. how do generative language models answer opinion polls?\.Sociological Methods & Research54\(3\),pp\. 1156–1196\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Buck and McFall \(2012\)N\. Buck and S\. McFallUnderstanding society: design overview\.Longitudinal and Life Course Studies3\(1\),pp\. 5–17\.Cited by:[Appendix C](https://arxiv.org/html/2608.19621#A3.SSx1.SSS0.Px2.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p5.1)\.
- Chenet al\.\(2025\)R\. Chen, A\. Arditi, H\. Sleight, O\. Evans, and J\. LindseyPersona vectors: monitoring and controlling character traits in language models\.arXiv preprint arXiv:2507\.21509\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Chunget al\.\(2023\)J\. J\. Y\. Chung, E\. Kamar, and S\. AmershiIncreasing diversity while maintaining accuracy: text data generation with large language models and human interventions\.InProceedings of the 61st annual meeting of the association for computational linguistics \(volume 1: Long papers\),pp\. 575–593\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Curran and Bauer \(2011\)P\. J\. Curran and D\. J\. BauerThe disaggregation of within\-person and between\-person effects in longitudinal models of change\.Annual review of psychology62\(1\),pp\. 583–619\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p3.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Duet al\.\(2026\)B\. Du, M\. Guo, S\. He, Z\. Ye, X\. Zhu, W\. Su, S\. Zhu, Y\. Zhou, Y\. Zhang, Q\. Ai, and Y\. LiuTwinVoice: a multi\-dimensional benchmark towards digital twins via LLM persona simulation\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 19604–19628\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.981),[Link](https://aclanthology.org/2026.findings-acl.981/)Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Duet al\.\(2025\)B\. Du, Z\. Ye, Z\. Wu, M\. A\. Jankowska, S\. Zhu, Q\. Ai, Y\. Zhou, and Y\. LiuSimVBG: simulating individual values by backstory generation\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 13104–13133\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx3.SSS0.Px1.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Elder Jr \(1998\)G\. H\. Elder JrThe life course as developmental theory\.Child development69\(1\),pp\. 1–12\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p3.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Geet al\.\(2024\)T\. Ge, X\. Chan, X\. Wang, D\. Yu, H\. Mi, and D\. YuScaling synthetic data creation with 1,000,000,000 personas\.arXiv preprint arXiv:2406\.20094\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Grattafioriet al\.\(2024\)A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p5.1)\.
- Guoet al\.\(2026a\)M\. Guo, Q\. Jiao, Z\. Shi, Y\. Quan, B\. Zhang, D\. Li, L\. Che, W\. Xu, S\. Liu, Z\. Liu, M\. Kapadia, V\. Pavlovic, J\. Liu, M\. Wang, Y\. Shi, D\. N\. Metaxas, and R\. TangMemEye: a visual\-centric evaluation framework for multimodal agent memory\.External Links:2605\.15128,[Document](https://dx.doi.org/10.48550/arXiv.2605.15128),[Link](https://arxiv.org/abs/2605.15128)Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Guoet al\.\(2026b\)M\. Guo, Z\. Ye, W\. Xu, X\. Zhu, W\. Hua, and D\. N\. MetaxasIndividual turing test: a case study of LLM\-based simulation using longitudinal personal data\.InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval,New York, NY, USA,pp\. 3770–3775\.External Links:[Document](https://dx.doi.org/10.1145/3805712.3809868),[Link](https://doi.org/10.1145/3805712.3809868)Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Haerpferet al\.\(2022\)C\. Haerpfer, R\. Inglehart, A\. Moreno, C\. Welzel, K\. Kizilova, J\. Diez\-Medrano, M\. Lagos, P\. Norris, E\. Ponarin, and B\. PuranenWorld values survey wave 7 \(2017\-2022\) cross\-national data\-set\.World Values Survey Association\.External Links:[Document](https://dx.doi.org/10.14281/18241.24)Cited by:[Appendix B](https://arxiv.org/html/2608.19621#A2.SSx1.SSS0.Px1.p1.1),[Figure 1](https://arxiv.org/html/2608.19621#Sx1.F1)\.
- Harriset al\.\(2019\)K\. M\. Harris, C\. T\. Halpern, E\. A\. Whitsel, J\. M\. Hussey, L\. A\. Killeya\-Jones, J\. Tabor, and S\. C\. DeanCohort profile: the national longitudinal study of adolescent to adult health \(add health\)\.International journal of epidemiology48\(5\),pp\. 1415–1415k\.Cited by:[Appendix C](https://arxiv.org/html/2608.19621#A3.SSx1.SSS0.Px1.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p5.1)\.
- Holtzmanet al\.\(2019\)A\. Holtzman, J\. Buys, L\. Du, M\. Forbes, and Y\. ChoiThe curious case of neural text degeneration\.arXiv preprint arXiv:1904\.09751\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Hu and Collier \(2024\)T\. Hu and N\. CollierQuantifying the persona effect in llm simulations\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 10289–10307\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px1.p1.1),[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px2.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Huet al\.\(2025\)Z\. Hu, J\. Lian, Z\. Xiao, M\. Xiong, Y\. Lei, T\. Wang, K\. Ding, Z\. Xiao, N\. J\. Yuan, and X\. XiePopulation\-aligned persona generation for llm\-based social simulation\.arXiv preprint arXiv:2509\.10127\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Kumaranet al\.\(2016\)D\. Kumaran, D\. Hassabis, and J\. L\. McClellandWhat learning systems do intelligent agents need? complementary learning systems theory updated\.Trends in Cognitive Sciences20\(7\),pp\. 512–534\.External Links:[Document](https://dx.doi.org/10.1016/j.tics.2016.05.004)Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p4.1)\.
- Lewiset al\.\(2020\)P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel,et al\.Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.Advances in neural information processing systems33,pp\. 9459–9474\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx3.SSS0.Px3.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Liet al\.\(2016\)J\. Li, M\. Galley, C\. Brockett, J\. Gao, and W\. B\. DolanA diversity\-promoting objective function for neural conversation models\.InProceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies,pp\. 110–119\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Li and Conrad \(2026\)M\. Li and F\. G\. ConradPersona\-based simulation of human opinion at population scale\.arXiv preprint arXiv:2603\.27056\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Liuet al\.\(2026\)A\. H\. Liu, K\. Khandelwal, S\. Subramanian, V\. Jouault, A\. Rastogi, A\. Sadé, A\. Jeffares, A\. Jiang, A\. Cahill, A\. Gavaudan,et al\.Ministral 3\.arXiv preprint arXiv:2601\.08584\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p5.1)\.
- Longet al\.\(2026\)L\. Long, Y\. He, W\. Ye, Y\. Pan, Y\. Lin, H\. Li, J\. Zhao, and W\. LiSeeing, listening, remembering, and reasoning: a multimodal agent with long\-term memory\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=PMz29A7Muq)Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Lutzet al\.\(2025\)M\. Lutz, I\. Sen, G\. Ahnert, E\. Rogers, and M\. StrohmaierThe prompt makes the person \(a\): a systematic evaluation of sociodemographic persona prompting for large language models\.arXiv preprint arXiv:2507\.16076\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Maharanaet al\.\(2024\)A\. Maharana, D\. Lee, S\. Tulyakov, M\. Bansal, F\. Barbieri, and Y\. FangEvaluating very long\-term conversational memory of llm agents\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 13851–13870\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx3.SSS0.Px2.p1.1),[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx3.SSS0.Px3.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- McClellandet al\.\(1995\)J\. L\. McClelland, B\. L\. McNaughton, and R\. C\. O’ReillyWhy there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory\.\.Psychological review102\(3\),pp\. 419\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p4.1),[LifeMem Overview](https://arxiv.org/html/2608.19621#Sx3.SSx2.p1.1)\.
- O’Reillyet al\.\(2014\)R\. C\. O’Reilly, R\. Bhattacharyya, M\. D\. Howard, and N\. KetzComplementary learning systems\.Cognitive Science38\(6\),pp\. 1229–1248\.External Links:[Document](https://dx.doi.org/10.1111/j.1551-6709.2011.01214.x)Cited by:[LifeMem Overview](https://arxiv.org/html/2608.19621#Sx3.SSx2.p1.1)\.
- O’Reilly and Norman \(2002\)R\. C\. O’Reilly and K\. A\. NormanHippocampal and neocortical contributions to memory: advances in the complementary learning systems framework\.Trends in Cognitive Sciences6\(12\),pp\. 505–510\.External Links:[Document](https://dx.doi.org/10.1016/S1364-6613%2802%2902005-3)Cited by:[LifeMem Overview](https://arxiv.org/html/2608.19621#Sx3.SSx2.p1.1)\.
- Parket al\.\(2023\)J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. BernsteinGenerative agents: interactive simulacra of human behavior\.InProceedings of the 36th annual acm symposium on user interface software and technology,pp\. 1–22\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Parket al\.\(2024\)J\. S\. Park, C\. Q\. Zou, A\. Shaw, B\. M\. Hill, C\. Cai, M\. R\. Morris, R\. Willer, P\. Liang, and M\. S\. BernsteinGenerative agent simulations of 1,000 people\.arXiv preprint arXiv:2411\.1010952\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Plattet al\.\(1999\)J\. Plattet al\.Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods\.Advances in large margin classifiers10\(3\),pp\. 61–74\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Prentice and Miller \(2007\)D\. A\. Prentice and D\. T\. MillerPsychological essentialism of human categories\.Current directions in psychological science16\(4\),pp\. 202–206\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Qwen Team \(2026\)Qwen TeamQwen3\.5: towards native multimodal agents\.Note:https://qwen\.ai/blog?id=qwen3\.5Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p5.1)\.
- Santurkaret al\.\(2023\)S\. Santurkar, E\. Durmus, F\. Ladhak, C\. Lee, P\. Liang, and T\. HashimotoWhose opinions do language models reflect?\.InInternational conference on machine learning,pp\. 29971–30004\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px1.p1.1),[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx1.SSS0.Px2.p1.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Sivakumaret al\.\(2025\)N\. Sivakumar, N\. Mackraz, S\. Khorshidi, K\. Patel, B\. Theobald, L\. Zappella, and N\. ApostoloffBias after prompting: persistent discrimination in large language models\.arXiv preprint arXiv:2509\.08146\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx2.SSS0.Px2.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Suet al\.\(2025\)W\. Su, Y\. Tang, Q\. Ai, J\. Yan, C\. Wang, H\. Wang, Z\. Ye, Y\. Zhou, and Y\. LiuParametric retrieval augmented generation\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1240–1250\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Vijayakumaret al\.\(2018\)A\. Vijayakumar, M\. Cogswell, R\. Selvaraju, Q\. Sun, S\. Lee, D\. Crandall, and D\. BatraDiverse beam search for improved description of complex scenes\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.32\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Wanget al\.\(2025a\)A\. Wang, J\. Morgenstern, and J\. P\. DickersonLarge language models that replace human participants can harmfully misportray and flatten identity groups\.Nature Machine Intelligence7\(3\),pp\. 400–411\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Wanget al\.\(2026\)H\. Wang, Y\. Zhou, B\. Du, Q\. Ai, and Y\. LiuParametric social identity injection and diversification in public opinion simulation\.arXiv preprint arXiv:2603\.16142\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Introduction](https://arxiv.org/html/2608.19621#Sx1.p3.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Wanget al\.\(2025b\)Q\. Wang, S\. Pan, T\. Linzen, and E\. BlackMultilingual prompting for improving llm generation diversity\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 6378–6400\.Cited by:[Appendix E](https://arxiv.org/html/2608.19621#A5.SSx2.SSS0.Px1.p1.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1),[Baselines](https://arxiv.org/html/2608.19621#Sx4.SSx2.p1.1)\.
- Wonget al\.\(2026\)J\. Wong, Y\. Orlovskiy, A\. Shypula, M\. Luo, S\. Seshia, and J\. GonzalezSimplestrat: diversifying language model generation with stratification\.Advances in Neural Information Processing Systems38,pp\. 116906–116935\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Xieet al\.\(2026\)Y\. Xie, L\. Liang, S\. Li, Y\. Lu, Z\. Xiao, M\. Shi, J\. Huang, M\. Wang, and Y\. XieEvaluating the statistical realism of llm\-generated social science data\.Proceedings of the National Academy of Sciences123\(19\),pp\. e2538145123\.Cited by:[Introduction](https://arxiv.org/html/2608.19621#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.19621#Sx2.p1.1)\.
- Zhanget al\.\(2025\)L\. H\. Zhang, S\. Milli, K\. Jusko, J\. Smith, B\. Amos, W\. Bouaziz, M\. Revel, J\. Kussman, Y\. Sheynin, L\. Titus,et al\.Cultivating pluralism in algorithmic monoculture: the community alignment dataset\.arXiv preprint arXiv:2507\.09650\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.
- Zhonget al\.\(2024\)W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. WangMemorybank: enhancing large language models with long\-term memory\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 19724–19731\.Cited by:[Related Work](https://arxiv.org/html/2608.19621#Sx2.p2.1)\.

## Appendix ALimitations

LifeMem is limited by the coverage and granularity of longitudinal surveys, which cannot fully capture an individual’s life experiences\. Surveys may omit events outside the questionnaire, record them only at coarse intervals, and fail to capture event duration or respondents’ subjective interpretations\. Thus, reconstructed trajectories remain partial representations of personal development rather than complete life histories\. Moreover, self\-reported survey responses may be affected by recall bias and question framing, and do not necessarily correspond to real\-world behavior\.

## Appendix BDetails of the Motivating Analyses

### WVS Static\-Profile Experiment

#### Experimental setup\.

We use Wave 7 of the World Values Survey \(WVS\)\([19](https://arxiv.org/html/2608.19621#bib.bib47)\)and randomly sample 2,000 respondents with seed 42\. For each respondent, we construct a static profile from answers in the WVS demographic section, corresponding to Questions Q260–Q290\. The profile is provided toLlama\-3\.1\-8B\-Instruct, which is instructed to simulate the corresponding respondent and answer the survey questions\. This setup represents a typical profile\-conditioned social simulation in which each agent is reconstructed from demographic attributes\.

#### Prompt template\.

The following prompt template is instantiated with the recoded demographic attributes of each respondent:

Prompt TemplateForget you are an AI model\. Simulate a human being\. Please answer the following question truthfully\.Please answer based on the following personal profile:You are <gender\>\. You are aged <age\>\. You live in <country\>\. You are <citizenship status\>\. You were born <birthplace status\>\. Your mother was born <mother’s birthplace status\>\. Your father was born <father’s birthplace status\>\. You live in <urbanicity\>, <settlement type\>, with a population of <population size\>\. Your household consists of <household size\> members\. You <parent cohabitation status\>\. You speak <home language\> at home\. You are <marital status\> and <children description\>\. Your education level is <education level\>\. Your spouse’s education level is <spouse education level\>\. You are <employment status\>\. You work as <occupation\>\. You work in <employment sector\>\. Your spouse is <spouse employment status\>\. Your spouse works as <spouse occupation\>\. You are <chief wage earner status\>\. During the past year, your family <household financial change\>\. You belong to <subjective social class\>\. Your income is in income group <income group\> \(1 = lowest, 10 = highest\)\. You identify as <religion\>\. Your ethnic group is <ethnic group\>\.Question:<question\>Options:<options\>Please ONLY output a number\. Do NOT write any words, symbols, punctuation, or explanations\.

#### Response processing\.

For both human respondents and their corresponding agents, the survey answers are converted into response vectors\. Each question dimension is standardized before dimensionality reduction, after which PCA is applied to obtain two\-dimensional representations\. Human and agent responses follow the same response labeling, processing, standardization, PCA, and visualization procedure\.

#### Socioeconomic status construction\.

We construct a composite socioeconomic status \(SES\) measure from four WVS variables: subjective social class \(Q287\), household economic situation \(Q286\), educational attainment \(Q275\), and income group \(Q288\)\.

Each variable is converted to a directionally consistent numerical scale such that larger values indicate higher socioeconomic status\. Letx~i​m\\widetilde\{x\}\_\{im\}denote the recoded value of variablemmfor respondentii, and letℳi\\mathcal\{M\}\_\{i\}denote the set of available SES variables for that respondent\. The composite score is

SESi=1\|ℳi\|​∑m∈ℳix~i​m\.\\operatorname\{SES\}\_\{i\}=\\frac\{1\}\{\|\\mathcal\{M\}\_\{i\}\|\}\\sum\_\{m\\in\\mathcal\{M\}\_\{i\}\}\\widetilde\{x\}\_\{im\}\.\(11\)
Respondents are divided into low\-, middle\-, and high\-SES groups according to the tertiles of the composite score\.

#### Silhouette score\.

We use the silhouette score to quantify how strongly the response representations separate according to the predefined SES groups\. LetCiC\_\{i\}denote the SES group containing individualii\. The average distance betweeniiand other members of the same group is

a⁡\(i\)=1\|Ci\|−1​∑j∈Cij≠id⁡\(i,j\)\.a\(i\)=\\frac\{1\}\{\|C\_\{i\}\|\-1\}\\sum\_\{\\begin\{subarray\}\{c\}j\\in C\_\{i\}\\\\ j\\neq i\\end\{subarray\}\}d\(i,j\)\.\(12\)
The minimum average distance betweeniiand any other SES group is

b⁡\(i\)=minC≠Ci⁡1\|C\|​∑j∈Cd⁡\(i,j\)\.b\(i\)=\\min\_\{C\\neq C\_\{i\}\}\\frac\{1\}\{\|C\|\}\\sum\_\{j\\in C\}d\(i,j\)\.\(13\)
The silhouette value for individualiiis

s⁡\(i\)=b⁡\(i\)−a⁡\(i\)max⁡\{a⁡\(i\),b⁡\(i\)\},s\(i\)=\\frac\{b\(i\)\-a\(i\)\}\{\\max\\\{a\(i\),b\(i\)\\\}\},\(14\)
and the overall silhouette score is

S=1N​∑i=1Ns⁡\(i\)\.S=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}s\(i\)\.\(15\)
Silhouette scores are computed in the standardized response space before PCA projection; PCA is used only for visualization\.

A score close to 1 indicates compact within\-group representations and clear between\-group separation, a score close to 0 indicates substantial overlap, and a negative score indicates that some individuals are closer to another group than to their assigned group\.

In this work, the silhouette score is used only to quantify the extent to which response representations align with predefined demographic groups\. It is not interpreted as a standalone measure of simulation quality\.

#### Result summary\.

Human responses are broadly dispersed and substantially overlap across SES groups, resulting in a silhouette score ofS=−0\.02S=\-0\.02\. In contrast, agent responses form more compact within\-group clusters and show clearer separation between SES groups, resulting inS=0\.19S=0\.19\. This pattern suggests that static demographic prompting may compress complex individuals into group\-typical representatives, thereby reducing within\-group heterogeneity and amplifying between\-group differences\. We treat this experiment as a motivating diagnostic that reveals a pattern consistent with identity essentialism in LLM agents\.

### Event RAG Retrieval\-Depth Analysis

We analyze how retrieval depth affects the effectiveness and computational cost of theEvent RAGbaseline\. Experiments are conducted on Add Health using Llama\-3\.1\-8B\-Instruct and 100 respondents sampled with random seed 42\. Event retrieval usesall\-MiniLM\-L6\-v2, consistent with the retrieval component used by LifeMem in this comparison\. Its lightweight architecture keeps retrieval overhead low while maintaining competitive retrieval quality\.

We vary the retrieval depth \(Top\-KK\) from 1 to 180\. For each value ofKK,Event RAGretrieves theKKhighest\-ranked life events relevant to each evaluation question and inserts them into the model prompt\. We report KL divergence, within\-group pairwise distance gap, and normalized entropy gap, where lower values indicate closer agreement with the corresponding human distributions\.

#### Runtime\.

Runtime is measured as the end\-to\-end wall\-clock duration of one completeEvent RAGexperiment over all 100 respondents and all six survey waves\. It therefore includes model loading, population initialization, life\-event loading, event retrieval, prompt construction, model inference, output parsing, and metric computation\.

#### Input tokens\.

For each retrieval depthKK, we average the estimated input tokens over all respondent–question prompts across all six survey waves\. The reported value therefore represents the mean input length per evaluation prompt\.

#### Comparison with LifeMem\.

Figure 3 in the main paper additionally marks LifeMem atK=5K=5as a reference\. While increasingEvent RAGretrieval depth initially improves distributional and diversity alignment, the gains largely plateau aroundK=40K=40and sometimes deteriorate, whereas runtime and input\-token use continue to increase\. Notably, LifeMem already achieves substantially better performance with onlyK=5K=5, suggesting that its gains are unlikely to be explained solely by retrieving more life events\. Instead, these results motivate combining selective retrieval with persistent parametric memory\.

## Appendix CDataset Construction and Preprocessing

### Dataset Overviews

Table 4:Number of respondents in the public\-use Add Health files across Waves I–VI\.Table 5:Number of respondents in the public\-use Understanding Society files across Waves 1–15\.#### Add Health\.

The National Longitudinal Study of Adolescent to Adult Health \(Add Health\) follows a nationally representative U\.S\. cohort initially recruited from students in Grades 7–12 during the 1994–1995 school year\([20](https://arxiv.org/html/2608.19621#bib.bib15)\)\. We use all six available survey waves: Wave I \(1994–1995\), Wave II \(1996\), Wave III \(2001–2002\), Wave IV \(2008–2009\), Wave V \(2016–2018\), and Wave VI \(2022–2025\)\. These waves trace respondents from adolescence through young adulthood and into early midlife\.

The survey content evolves with the cohort’s life stage\. Waves I–II focus on adolescent health, education, family and peer relationships, and risk behavior\. Wave III expands to relationship, fertility, education, and employment histories during the transition to adulthood\. Waves IV–VI further examine adult health, socioeconomic circumstances, cognition, caregiving, and family life\. Together, these waves provide rich longitudinal records spanning education, employment, health, family, and residential transitions\.

The public\-use files contain between 3,937 and 6,504 respondents per wave across Waves I–VI\. Detailed wave\-level sample sizes are reported in Table[4](https://arxiv.org/html/2608.19621#A3.T4)\. Respondents are linked across waves using the stable person identifierAID\.

#### Understanding Society\.

Understanding Society, also known as the UK Household Longitudinal Study \(UKHLS\), is an annual household panel survey that began in 2009\([8](https://arxiv.org/html/2608.19621#bib.bib14)\)\. We use UKHLS Waves 1–15, covering interviews conducted from 2009 to 2024\. Each wave is fielded over an overlapping period, while individual sample members are generally interviewed approximately one year apart\.

Understanding Society follows members of sampled UK households and covers household composition, family relationships, education, employment, income, housing, health and well\-being, finances, political attitudes, and social participation\. Its annual panel structure and broad question coverage across fifteen waves support the evaluation of changes in respondents’ answers between adjacent waves\.

The public\-use files contain between 41,601 and 77,495 respondents per wave across Waves 1–15\. Detailed wave\-level sample sizes are reported in Table[5](https://arxiv.org/html/2608.19621#A3.T5)\. Respondents are linked across waves using the stable person identifierpidp\.

### Respondent Alignment and Sampling

To construct complete longitudinal trajectories, we align respondents using stable cross\-wave identifiers and retain only individuals observed in every wave of the corresponding dataset\. After alignment, 2,048 Add Health respondents are observed with the sameAIDacross all six waves, while 14,104 Understanding Society respondents are observed with the samepidpacross all fifteen waves\.

For the main experiments, we randomly sample 100 respondents from each aligned pool using seed 42\. The sampled respondent identifiers are fixed and shared across all methods and backbone models within each dataset\. Consequently, all comparisons use the same individuals, survey questions, and observed life histories, thereby controlling for differences in respondent composition\. Additional experiments using seeds 43 and 44 evaluate robustness to respondent sampling, as reported in Appendix[H](https://arxiv.org/html/2608.19621#A8.SSx1)\.

### Variable Classification

#### Structured Question Records\.

Each selected variable is represented as a structured question record with the following fields:

- •Variable: the survey variable name\.
- •Section: the questionnaire section\.
- •Question: the natural\-language survey question\.
- •Options: the response options for closed\-ended questions\.
- •Note: additional constraints for open\-ended responses\.

#### Functional Categories\.

We organize survey variables into three functional categories:demographic,life event, andevaluate\. Thedemographicvariables are used to construct the initial profile of an agent, including background and profile attributes such as age, gender, education, employment, household composition, region, religion, and family background\. Thelife eventvariables describe longitudinal experiences and state changes, such as schooling, work transitions, partnership changes, fertility, housing changes, health events, caregiving, economic hardship, and other major events\. These variables are used to construct structured memory and parametric memory updates across waves\. Theevaluatevariables are reserved exclusively for model evaluation and are excluded from profile construction, structured memory, and parametric\-memory updates\. Examples include questions about residential preferences, internet use, social participation, political attitudes, and subjective well\-being\.

Table 6:Numbers of demographic, life\-event, and evaluation variables selected for each dataset and survey wave\. Demo\., Event, and Eval\. denote demographic, life\-event, and evaluation variables, respectively\. UKHLS denotes Understanding Society\.
#### Classification and Filtering\.

Variables are classified using a rule\-based semantic matching procedure followed by response\-quality filtering\. The semantic classifier considers the section, question, options, note, and variable fields\. We maintain category\-specific keyword dictionaries for the three categories\. After semantic matching, we filter variables using response\-quality statistics computed from the human answer files\. For each variable, we compute the total number of responses, the number of valid responses, and the number and rate of missing responses\. Variables with severe missingness or too few valid responses are removed\. Table[6](https://arxiv.org/html/2608.19621#A3.T6)reports the number of selected variables in each category for every dataset and wave\.

#### Examples of Classified Variables\.

The following examples illustrate the three variable categories using questions from Add Health Wave 1\.

Demographic VariablesVariable:BIO\_SEXSection:Section A: Setup of CAPI InterviewQuestion:What is your gender?Options: 1: Male; 2: FemaleVariable:H1GI1MSection:Section 1: General IntroductoryQuestion:Which month were you born in?Options: 1: January; 2: February; 3: March; 4: April; 5: May; 6: June; 7: July; 8: August; 9: September; 10: October; 11: November; 12: December

Life\-Event VariablesVariable:H1GH2Section:Section 3: General HealthQuestion:How often have you had a headache?Options: 0: Never; 1: Just a few times; 2: About once a week; 3: Almost every day; 4: Every dayVariable:H1GH43Section:Section 3: General HealthQuestion:During the past 30 days, how often did you drive a car or other vehicle when you had been drinking alcohol?Options: 0: Never; 1: 1 time; 2: 2 or 3 times; 3: 4 or 5 times; 4: 6 or more times

Evaluation VariablesVariable:H1ED19Section:Section 5: Academics and EducationQuestion:You feel close to people at your school\.Options: 1: Strongly agree; 2: Agree; 3: Neither agree nor disagree; 4: Disagree; 5: Strongly disagreeVariable:H1PF4Section:Section 18: Personality and FamilyQuestion:You are satisfied with the way your mother and you communicate with each other\.Options: 1: Strongly agree; 2: Agree; 3: Neither agree nor disagree; 4: Disagree; 5: Strongly disagree

### Common Longitudinal Evaluation Set

In addition to the wave\-specific evaluation variables, we construct a common longitudinal evaluation set for Understanding Society\. This set provides a fixed target space across waves, enabling us to evaluate whether agents reproduce not only population\-level response distributions but also changes in the same survey variables over time\.

We first retain variables that appear in the question files for all 15 waves, yielding 198 shared variables\. We then exclude variables used as demographic or life\-event inputs, leaving 136 candidates\. After semantic and response\-quality filtering, the final evaluation set contains 50 variables covering housing, employment, income, caregiving, family, education, and related life outcomes\.

Transition\-distribution JS divergence is computed only on Understanding Society because it provides a fixed set of comparable evaluation variables across all 15 waves\. Add Health contains fewer waves and substantially less overlap among comparable evaluation variables, so we do not report Transition JS for Add Health\.

## Appendix DShared Experimental Setup

### Inference Protocol

We evaluate all methods using three instruction\-tuned backbone models: Llama\-3\.1\-8B\-Instruct, Ministral\-3\-8B\-Instruct\-2512, and Qwen3\.5\-9B\. Thinking mode is disabled for Qwen3\.5\-9B, and each model uses its native tokenizer and chat template\. All methods share the same deterministic decoding configuration, with temperature00, sampling disabled, a maximum input length of 4,096 tokens, and a maximum output length of 16 tokens\. Inference is conducted with a batch size of four usingbfloat16precision\.

All methods also share the same survey\-response prompt structure\. Method\-specific context, such as a demographic profile or retrieved life events, is inserted before the question when applicable\. For closed\-ended questions, the model is required to return exactly one option number\. For open\-ended questions, it is instead instructed to provide a concise response without explanation\.

Shared Survey PromptYou are not an AI assistant; you are role\-playing a human survey respondent\.Question:<question\>For closed\-ended questions:Options:<options\>Answer with exactly one option number\. Do not provide any explanation\.For open\-ended questions:Notes:<notes\>Answer concisely\. Do not provide any explanation\.

### Survey\-to\-Statement Conversion

We convert each survey question–answer pair into one second\-person statement and one first\-person statement\. The second\-person statements are used to construct demographic profiles and structured life\-event memory, while the first\-person statements serve as reconstruction targets for parametric\-memory training\.

Conversion is performed using gpt\-3\.5\-turbo\-0125, with temperature00and a maximum output length of 2,048 tokens\. The conversion prompt is shown below\.

Survey\-to\-Statement Conversion PromptYou convert survey question\-answer pairs into concise personal statements\.Rewrite the provided question and answer as one faithful second\-person statement and one first\-person statement about the respondent\.Rules:1\. Preserve the exact meaning of the answer\.2\. Do not add information, explanations, causes, stereotypes, or assumptions\.3\. Preserve negation, quantities, dates, frequencies, and time ranges\.4\. For life events, do not convert an event into a permanent trait unless the answer says so\.5\. Return valid JSON only with keyssecond\_personandfirst\_person\.Input:<input\>

For example, the survey response

Example Survey ResponseQuestion:How often have you had trouble falling asleep or staying asleep?Answer:Never

is converted into:

Example Converted StatementsSecond\_Person:You have never had trouble falling asleep or staying asleep\.First\_Person:I have never had trouble falling asleep or staying asleep\.

This conversion preserves the observed survey response while producing textual representations suitable for profile construction, event retrieval, and adapter training\.

## Appendix EBaseline Implementation Details

### Static Conditioning

#### Direct\.

Direct\([3](https://arxiv.org/html/2608.19621#bib.bib2);[40](https://arxiv.org/html/2608.19621#bib.bib3);[6](https://arxiv.org/html/2608.19621#bib.bib4);[22](https://arxiv.org/html/2608.19621#bib.bib5)\)receives only the shared survey instruction, the current question, its response options or notes, and the corresponding output constraint\. It does not receive demographic information or longitudinal life events\.

#### Profile\.

Profile\([3](https://arxiv.org/html/2608.19621#bib.bib2);[40](https://arxiv.org/html/2608.19621#bib.bib3);[6](https://arxiv.org/html/2608.19621#bib.bib4);[22](https://arxiv.org/html/2608.19621#bib.bib5)\)additionally conditions the model on each respondent’s fixed initial demographic profile\.

Profile Conditioning PromptYou are answering a longitudinal social survey as the described person\.Demographic profile:<profile\_text\>

### Diversity\-Oriented Prompting

#### Multilingual\.

Multilingual\([46](https://arxiv.org/html/2608.19621#bib.bib19)\)follows theDirectsetting but translates the complete survey prompt into four widely used languages: Chinese, Spanish, English, and Arabic\. For each respondent\-question pair, one of the four languages is selected uniformly at random using a deterministic hash of the global seed, wave index, respondent ID, and question variable\.

#### Anti\-Stereotype\.

Anti\-Stereotype\([41](https://arxiv.org/html/2608.19621#bib.bib20)\)follows theProfilesetting and appends an explicit instruction discouraging demographic stereotyping\. This baseline evaluates whether explicit anti\-stereotyping guidance can mitigate demographic overgeneralization\.

Anti\-Stereotype InstructionDo not assume that one demographic attribute determines another\. When information is missing, preserve uncertainty rather than filling it with stereotypes\.

### Non\-Parametric Memory

#### SimVBG\.

SimVBG\([13](https://arxiv.org/html/2608.19621#bib.bib11)\)first uses the current backbone model to expand each respondent’s demographic profile into a comprehensive second\-person background story, which remains fixed across all survey waves\. Story generation uses a batch size of four, temperature00, sampling disabled, and a maximum output length of 2,048 tokens\.

The generation prompt includes the following instructions:

Background\-Story Generation PromptYou are a background story writer\. Your task is to craft a comprehensive backstory for a person based on the information provided below\. You must include every single data point from the original information, without exception\.This person’s information:<profile\_text\>

During evaluation, the generated story is inserted before the shared survey prompt:

SimVBG Evaluation PromptYou are answering a longitudinal social survey as the person described in the background story\.Background story:<background\_story\>

#### Full History\.

Full History\([31](https://arxiv.org/html/2608.19621#bib.bib21)\)conditions the model on the demographic profile and all respondent\-specific life events accumulated through the current wave:

Full\-History PromptYou are answering a longitudinal social survey as the described person\.Demographic profile:<profile\_text\>Life history:–<event\_1\>–<event\_2\>…

When the accumulated history exceeds the available input budget, events are retained in reverse chronological order under a character budget of3×max\_input\_length3\\times\\texttt\{max\\\_input\\\_length\}\. Recent events are therefore preserved first, while older events may be truncated\.

#### Event RAG\.

Event RAG\([25](https://arxiv.org/html/2608.19621#bib.bib22);[31](https://arxiv.org/html/2608.19621#bib.bib21)\)conditions the model on the demographic profile and the five accumulated life events most relevant to the current evaluation question\. It usesbge\-m3to encode each evaluation question and each life\-event text, where a life\-event text concatenates the wave index, survey section, original question, answer text, and second\-person statement\.

For an eventei,τe\_\{i,\\tau\}observed for respondentiiat waveτ\\tauand a questionqtq\_\{t\}asked at wavett, the retrieval score is

s⁡\(qt,ei,τ\)=sim⁡\(qt,ei,τ\)⋅αrett−τ,s\(q\_\{t\},e\_\{i,\\tau\}\)=\\operatorname\{sim\}\(q\_\{t\},e\_\{i,\\tau\}\)\\cdot\\alpha\_\{\\mathrm\{ret\}\}^\{\\,t\-\\tau\},\(16\)whereαret=0\.9\\alpha\_\{\\mathrm\{ret\}\}=0\.9controls recency decay\. The semantic similarity is computed as

sim⁡\(qt,ei,τ\)=𝐡qt⊤​𝐡ei,τ,\\operatorname\{sim\}\(q\_\{t\},e\_\{i,\\tau\}\)=\\mathbf\{h\}\_\{q\_\{t\}\}^\{\\top\}\\mathbf\{h\}\_\{e\_\{i,\\tau\}\},\(17\)where𝐡qt\\mathbf\{h\}\_\{q\_\{t\}\}and𝐡ei,τ\\mathbf\{h\}\_\{e\_\{i,\\tau\}\}areℓ2\\ell\_\{2\}\-normalized embeddings frombge\-m3; hence the dot product is cosine similarity\. The five highest\-scoring events are inserted into the prompt\.

Event RAG PromptYou are answering a longitudinal social survey as the described person\.Demographic profile:<profile\_text\>Relevant life experiences:–<retrieved\_event\_1\>–<retrieved\_event\_2\>…

### Control Baseline

#### Random Event\.

Random Eventuses the same prompt structure and the same number of events asEvent RAG\. Instead of retrieving respondent\-specific events, it samples five events from other agents using a pseudorandom generator initialized with the global simulation seed\. This baseline preserves the amount and format of event information while removing the correspondence between the supplied events and the target respondent\.

## Appendix FLifeMem Implementation Details

LifeMem combines demographic profile conditioning, structured life\-event memory, and agent\-specific parametric memory\. At each survey wave, the system updates and evaluates every agent sequentially\.

For each agent at every wave, LifeMem first loads the newly observed life events and updates the agent\-specific LoRA adapter using both current and replayed events\. It then adds the current events to structured memory, retrieves relevant events for each evaluation question, and constructs the corresponding prompt\. After activating the agent\-specific adapter, the model generates responses in batches, which are subsequently parsed and evaluated\. Finally, the updated LoRA state is saved for use in the next wave\.

### Structured Life\-Event Memory

LifeMem usesall\-MiniLM\-L6\-v2as a frozen retrieval encoder\. Its lightweight architecture provides a practical balance between semantic retrieval quality and computational efficiency\. Apart from the encoder choice, its retrieval procedure, including top\-KK, similarity scoring, and recency weighting, is identical to that ofEvent RAG\. The recency term favors recent experiences while allowing older events to be retrieved when their semantic relevance remains high\.

### Agent\-Specific Parametric Memory

Each agent maintains an independent LoRA adapter, while the parameters of the underlying language model remain frozen\. Table[7](https://arxiv.org/html/2608.19621#A6.T7)reports the default adapter and optimization settings\.

Table 7:Default configuration for agent\-specific LoRA updates\.Algorithm 1Agent\-specific parametric\-memory update1:foreach wave

ttdo

2:foreach agent

iido

3:

ℰi,tcur←\\mathcal\{E\}\_\{i,t\}^\{\\mathrm\{cur\}\}\\leftarrowevents of agent

iiobserved at wave

tt
4:

ℰi,trep←\\mathcal\{E\}\_\{i,t\}^\{\\mathrm\{rep\}\}\\leftarrowsample

RRevents from agent

ii’s previous events

5:

ℬi,t←∅\\mathcal\{B\}\_\{i,t\}\\leftarrow\\emptyset
6:foreach event

e∈ℰi,tcure\\in\\mathcal\{E\}\_\{i,t\}^\{\\mathrm\{cur\}\}do

7:Add

\(e\.question,e\.answer\)\(e\.\\mathrm\{question\},e\.\\mathrm\{answer\}\)to

ℬi,t\\mathcal\{B\}\_\{i,t\}
8:Add

\(e\.statement,e\.first\_person\)\(e\.\\mathrm\{statement\},e\.\\mathrm\{first\\\_person\}\)to

ℬi,t\\mathcal\{B\}\_\{i,t\}
9:endfor

10:foreach event

e∈ℰi,trepe\\in\\mathcal\{E\}\_\{i,t\}^\{\\mathrm\{rep\}\}do

11:Add

\(e\.question,e\.answer\)\(e\.\\mathrm\{question\},e\.\\mathrm\{answer\}\)to

ℬi,t\\mathcal\{B\}\_\{i,t\}
12:Add

\(e\.statement,e\.first\_person\)\(e\.\\mathrm\{statement\},e\.\\mathrm\{first\\\_person\}\)to

ℬi,t\\mathcal\{B\}\_\{i,t\}
13:Mark both examples with

replay=true\\mathrm\{replay\}=\\mathrm\{true\}
14:endfor

15:Activate agent

ii’s LoRA adapter

16:Save the pre\-update parameters

Θi,tbefore←Θi,t\\Theta\_\{i,t\}^\{\\mathrm\{before\}\}\\leftarrow\\Theta\_\{i,t\}
17:foreach training epochdo

18:foreach minibatch

ℳ⊂ℬi,t\\mathcal\{M\}\\subset\\mathcal\{B\}\_\{i,t\}do

19:foreach example

z∈ℳz\\in\\mathcal\{M\}do

20:

ℓz←CE⁡\(fΘi,t​\(zinput\),zassistant\)\\ell\_\{z\}\\leftarrow\\operatorname\{CE\}\\\!\\left\(f\_\{\\Theta\_\{i,t\}\}\(z\_\{\\mathrm\{input\}\}\),z\_\{\\mathrm\{assistant\}\}\\right\)
21:if

zzis a replay examplethen

22:

ℓz←η​ℓz\\ell\_\{z\}\\leftarrow\\eta\\ell\_\{z\}
23:endif

24:endfor

25:

ℒCE←1\|ℳ\|​∑z∈ℳℓz\\mathcal\{L\}\_\{\\mathrm\{CE\}\}\\leftarrow\\frac\{1\}\{\|\\mathcal\{M\}\|\}\\sum\_\{z\\in\\mathcal\{M\}\}\\ell\_\{z\}
26:

ℒ←ℒCE\+γ​‖Θi,t−Θi,tbefore‖22\\mathcal\{L\}\\leftarrow\\mathcal\{L\}\_\{\\mathrm\{CE\}\}\+\\gamma\\left\\\|\\Theta\_\{i,t\}\-\\Theta\_\{i,t\}^\{\\mathrm\{before\}\}\\right\\\|\_\{2\}^\{2\}
27:Backpropagate

ℒ\\mathcal\{L\}
28:Clip gradients usingmax\_grad\_norm

29:Update the active adapter using the optimizer

30:endfor

31:endfor

32:Save agent

ii’s updated LoRA adapter state

33:endfor

34:endfor

#### Training examples\.

Each life event is converted into two training examples\. The first teaches the association between the original survey question and the respondent’s observed answer:

Question–Answer Training ExampleUser:<event\_question\>Assistant:<event\_answer\_text\>

The second reconstructs a first\-person self\-memory from the corresponding second\-person event statement:

Self\-Memory Reconstruction ExampleUser:Recall the following personal information: <second\_person\_event\>Assistant:<first\_person\_event\>

The adapter therefore learns both the mapping from survey questions to observed responses and the reconstruction of life events as first\-person personal memories\.

#### Sequential adapter updates\.

At each wave, LifeMem updates every agent’s adapter using the events newly observed at that wave together with replayed events sampled from the same agent’s history\. Algorithm[1](https://arxiv.org/html/2608.19621#alg1)summarizes the training procedure\.

Replay events are sampled exclusively from the same agent’s previously observed life events\. By default, we uniformly sample up toR=4R=4historical events and assign replay examples a loss weight ofη=0\.5\\eta=0\.5\. Each replayed event yields the same question–answer and self\-memory reconstruction examples as a current event\. Sampling uses a pseudorandom generator initialized with the global simulation seed\. When fewer thanRRhistorical events are available, all available events are included\.

Stability regularization is enabled by default\. Before updating an agent at wavett, we store the current adapter parametersΘi,tbefore\\Theta\_\{i,t\}^\{\\mathrm\{before\}\}and add

γ​‖Θi,t−Θi,tbefore‖22\\gamma\\left\\\|\\Theta\_\{i,t\}\-\\Theta\_\{i,t\}^\{\\mathrm\{before\}\}\\right\\\|\_\{2\}^\{2\}\(18\)
to the training objective, whereγ=0\.01\\gamma=0\.01\. This term limits abrupt parameter changes and reduces the extent to which newly observed events overwrite previously accumulated parametric memory\.

During inference, multiple evaluation questions for the same agent are generated in batches while that agent’s LoRA adapter remains active\. This avoids repeatedly switching adapters between individual questions and ensures that all responses for an agent at a given wave use the same parametric state\.

## Appendix GEvaluation Metrics

All metrics are computed only from valid human–model response pairs\. Specifically, a sample is retained only when the human response corresponds to a predefined valid option and the model output can be parsed into a valid option from the same response space\. Human responses coded as missing, refusal, “do not know,” or “not applicable” are excluded\. This shared filtering rule prevents differences in missing\-response behavior from affecting comparisons between methods\.

Let𝒞\\mathcal\{C\}denote the set of eligible wave–question cells\. For a cell\(t,q\)∈𝒞\(t,q\)\\in\\mathcal\{C\}, let𝒴t​q\\mathcal\{Y\}\_\{tq\}be its set of valid response labels, and letnht​q​\(y\)n\_\{h\}^\{tq\}\(y\)andnmt​q​\(y\)n\_\{m\}^\{tq\}\(y\)denote the numbers of valid human and model responses assigned to labelyy, respectively\.

### KL Divergence

For each wave–question cell, we construct empirical human and model response distributions\. To avoid zero probabilities, we addϵ=10−9\\epsilon=10^\{\-9\}to the count of every valid label:

Pht​q​\(y\)=nht​q​\(y\)\+ϵ∑y′∈𝒴t​qnht​q​\(y′\)\+\|𝒴t​q\|​ϵ,P\_\{h\}^\{tq\}\(y\)=\\frac\{n\_\{h\}^\{tq\}\(y\)\+\\epsilon\}\{\\sum\_\{y^\{\\prime\}\\in\\mathcal\{Y\}\_\{tq\}\}n\_\{h\}^\{tq\}\(y^\{\\prime\}\)\+\|\\mathcal\{Y\}\_\{tq\}\|\\epsilon\},\(19\)and

Pmt​q​\(y\)=nmt​q​\(y\)\+ϵ∑y′∈𝒴t​qnmt​q​\(y′\)\+\|𝒴t​q\|​ϵ\.P\_\{m\}^\{tq\}\(y\)=\\frac\{n\_\{m\}^\{tq\}\(y\)\+\\epsilon\}\{\\sum\_\{y^\{\\prime\}\\in\\mathcal\{Y\}\_\{tq\}\}n\_\{m\}^\{tq\}\(y^\{\\prime\}\)\+\|\\mathcal\{Y\}\_\{tq\}\|\\epsilon\}\.\(20\)
We then compute the KL divergence from the human distribution to the model distribution:

KLt​q=KL\(Pht​q∥Pmt​q\)=∑y∈𝒴t​qPht​q\(y\)logPht​q​\(y\)Pmt​q​\(y\)\.\\operatorname\{KL\}\_\{tq\}=\\operatorname\{KL\}\\left\(P\_\{h\}^\{tq\}\\,\\\|\\,P\_\{m\}^\{tq\}\\right\)=\\sum\_\{y\\in\\mathcal\{Y\}\_\{tq\}\}P\_\{h\}^\{tq\}\(y\)\\log\\frac\{P\_\{h\}^\{tq\}\(y\)\}\{P\_\{m\}^\{tq\}\(y\)\}\.\(21\)
The reported KL divergence is the unweighted mean across all eligible wave–question cells:

KL=1\|𝒞\|​∑\(t,q\)∈𝒞KLt​q\.\\operatorname\{KL\}=\\frac\{1\}\{\|\\mathcal\{C\}\|\}\\sum\_\{\(t,q\)\\in\\mathcal\{C\}\}\\operatorname\{KL\}\_\{tq\}\.\(22\)
Lower values indicate closer alignment between the model and human response distributions\.

### Normalized Entropy Gap

For a response distributionPPover the valid label set𝒴t​q\\mathcal\{Y\}\_\{tq\}, normalized entropy is defined as

H\(P\)=−1log⁡\|𝒴t​q\|∑y∈𝒴t​qP\(y\)logP\(y\)\.H\(P\)=\-\\frac\{1\}\{\\log\|\\mathcal\{Y\}\_\{tq\}\|\}\\sum\_\{y\\in\\mathcal\{Y\}\_\{tq\}\}P\(y\)\\log P\(y\)\.\(23\)The normalization bounds entropy between zero and one for questions with at least two valid response options\. We compute the human and model entropies separately:

Hht​q=H⁡\(Pht​q\),Hmt​q=H⁡\(Pmt​q\)\.H\_\{h\}^\{tq\}=H\\\!\\left\(P\_\{h\}^\{tq\}\\right\),\\qquad H\_\{m\}^\{tq\}=H\\\!\\left\(P\_\{m\}^\{tq\}\\right\)\.\(24\)
The entropy gap for each wave–question cell is the absolute difference between them:

EntropyGapt​q=\|Hmt​q−Hht​q\|\.\\operatorname\{EntropyGap\}\_\{tq\}=\\left\|H\_\{m\}^\{tq\}\-H\_\{h\}^\{tq\}\\right\|\.\(25\)
The reported normalized entropy gap is

NormEntropyGap=1\|𝒞\|​∑\(t,q\)∈𝒞EntropyGapt​q\.\\operatorname\{NormEntropyGap\}=\\frac\{1\}\{\|\\mathcal\{C\}\|\}\\sum\_\{\(t,q\)\\in\\mathcal\{C\}\}\\operatorname\{EntropyGap\}\_\{tq\}\.\(26\)
A smaller value indicates that the overall concentration or dispersion of model responses is closer to that of human responses\.

### Within\-Group Pairwise Distance Gap

Table 8:Demographic variables used to define groups for the within\-group pairwise distance metric\. UKHLS denotes Understanding Society\.The within\-group pairwise distance measures whether response diversity within demographic groups is preserved\. We form groups separately using each discrete demographic variable\. The grouping variables are shown in Table[8](https://arxiv.org/html/2608.19621#A7.T8)\.

Consider demographic variablegg, group valuevv, wavett, and questionqq\. Letnht​q​g​v​\(y\)n\_\{h\}^\{tqgv\}\(y\)be the number of human respondents in the group who select optionyy, and let

nht​q​g​v=∑y∈𝒴t​qnht​q​g​v​\(y\)\.n\_\{h\}^\{tqgv\}=\\sum\_\{y\\in\\mathcal\{Y\}\_\{tq\}\}n\_\{h\}^\{tqgv\}\(y\)\.\(27\)For groups containing at least two valid responses, the human within\-group categorical pairwise distance is

Dh​\(t,q,g,v\)=1−∑y∈𝒴t​qnht​q​g​v​\(y\)​\(nht​q​g​v​\(y\)−1\)nht​q​g​v​\(nht​q​g​v−1\)\.D\_\{h\}\(t,q,g,v\)=1\-\\frac\{\\sum\_\{y\\in\\mathcal\{Y\}\_\{tq\}\}n\_\{h\}^\{tqgv\}\(y\)\\left\(n\_\{h\}^\{tqgv\}\(y\)\-1\\right\)\}\{n\_\{h\}^\{tqgv\}\\left\(n\_\{h\}^\{tqgv\}\-1\\right\)\}\.\(28\)
This quantity is the probability that two distinct respondents sampled from the same group provide different answers\. The model distance is computed analogously:

Dm​\(t,q,g,v\)=1−∑y∈𝒴t​qnmt​q​g​v​\(y\)​\(nmt​q​g​v​\(y\)−1\)nmt​q​g​v​\(nmt​q​g​v−1\)\.D\_\{m\}\(t,q,g,v\)=1\-\\frac\{\\sum\_\{y\\in\\mathcal\{Y\}\_\{tq\}\}n\_\{m\}^\{tqgv\}\(y\)\\left\(n\_\{m\}^\{tqgv\}\(y\)\-1\\right\)\}\{n\_\{m\}^\{tqgv\}\\left\(n\_\{m\}^\{tqgv\}\-1\\right\)\}\.\(29\)
Group cells with fewer than two valid responses are excluded\. Let𝒢t​q\\mathcal\{G\}\_\{tq\}denote the set of eligible demographic\-variable–value cells for wavettand questionqq\. We first average the group\-specific distances:

Dh​\(t,q\)=1\|𝒢t​q\|​∑\(g,v\)∈𝒢t​qDh​\(t,q,g,v\),D\_\{h\}\(t,q\)=\\frac\{1\}\{\|\\mathcal\{G\}\_\{tq\}\|\}\\sum\_\{\(g,v\)\\in\\mathcal\{G\}\_\{tq\}\}D\_\{h\}\(t,q,g,v\),\(30\)Dm​\(t,q\)=1\|𝒢t​q\|​∑\(g,v\)∈𝒢t​qDm​\(t,q,g,v\)\.D\_\{m\}\(t,q\)=\\frac\{1\}\{\|\\mathcal\{G\}\_\{tq\}\|\}\\sum\_\{\(g,v\)\\in\\mathcal\{G\}\_\{tq\}\}D\_\{m\}\(t,q,g,v\)\.\(31\)
The wave–question\-level gap is then

WGGapt​q=\|Dm​\(t,q\)−Dh​\(t,q\)\|\.\\operatorname\{WGGap\}\_\{tq\}=\\left\|D\_\{m\}\(t,q\)\-D\_\{h\}\(t,q\)\\right\|\.\(32\)
Finally, the reported within\-group pairwise distance gap is

WGPairwiseDistGap=1\|𝒞WG\|​∑\(t,q\)∈𝒞WGWGGapt​q,\\operatorname\{WGPairwiseDistGap\}=\\frac\{1\}\{\|\\mathcal\{C\}\_\{\\mathrm\{WG\}\}\|\}\\sum\_\{\(t,q\)\\in\\mathcal\{C\}\_\{\\mathrm\{WG\}\}\}\\operatorname\{WGGap\}\_\{tq\},\(33\)where𝒞WG\\mathcal\{C\}\_\{\\mathrm\{WG\}\}contains wave–question cells with at least one eligible group cell\. Lower values indicate closer alignment between model and human within\-group response diversity\.

### Transition\-Distribution JS Divergence

Transition\-distribution JS divergence evaluates whether agents reproduce longitudinal patterns of response change rather than only the marginal response distribution at an individual wave\. It is computed using the common longitudinal evaluation set described in Appendix[C](https://arxiv.org/html/2608.19621#A3.SSx4)\.

For a common questionqq, respondentii, and two adjacent wavesttandt\+1t\+1, the human response transition is

yi,tq→yi,t\+1q,y\_\{i,t\}^\{q\}\\rightarrow y\_\{i,t\+1\}^\{q\},\(34\)while the corresponding model transition is

y^i,tq→y^i,t\+1q\.\\hat\{y\}\_\{i,t\}^\{q\}\\rightarrow\\hat\{y\}\_\{i,t\+1\}^\{q\}\.\(35\)
Table 9:Robustness across respondent samples on Add Health and Understanding Society\. Results report means across three independently sampled respondent sets generated with seeds 42, 43, and 44, with the corresponding±\\pmstandard deviations shown as subscripts\. Each sample contains 100 respondents per dataset\. KL Div\., WG Gap, and Ent\. Gap denote KL divergence, within\-group pairwise distance gap, and normalized entropy gap\. Lower values are better for all metrics\. Bold marks the best mean result for each model and metric; LifeMem is shaded light gray\.We pool valid transitions across respondents and adjacent\-wave pairs to construct empirical distributions over ordered response pairs:

Phq​\(a,b\)=Pr⁡\(ytq=a,yt\+1q=b\),P\_\{h\}^\{q\}\(a,b\)=\\Pr\\left\(y\_\{t\}^\{q\}=a,y\_\{t\+1\}^\{q\}=b\\right\),\(36\)Pmq​\(a,b\)=Pr⁡\(y^tq=a,y^t\+1q=b\),P\_\{m\}^\{q\}\(a,b\)=\\Pr\\left\(\\hat\{y\}\_\{t\}^\{q\}=a,\\hat\{y\}\_\{t\+1\}^\{q\}=b\\right\),\(37\)where\(a,b\)∈𝒴q×𝒴q\(a,b\)\\in\\mathcal\{Y\}\_\{q\}\\times\\mathcal\{Y\}\_\{q\}represents an ordered transition such as→11\\\!\\rightarrow\\\!1,→21\\\!\\rightarrow\\\!2, or→12\\\!\\rightarrow\\\!1\. A transition is included only when both adjacent human responses are valid and both corresponding model responses can be parsed into valid options\.

For each common question, we define the midpoint distribution

Mq=12​\(Phq\+Pmq\),M^\{q\}=\\frac\{1\}\{2\}\\left\(P\_\{h\}^\{q\}\+P\_\{m\}^\{q\}\\right\),\(38\)and compute

JStransq=12KL\(Phq∥Mq\)\+12KL\(Pmq∥Mq\)\.\\operatorname\{JS\}\_\{\\mathrm\{trans\}\}^\{q\}=\\frac\{1\}\{2\}\\operatorname\{KL\}\\left\(P\_\{h\}^\{q\}\\,\\\|\\,M^\{q\}\\right\)\+\\frac\{1\}\{2\}\\operatorname\{KL\}\\left\(P\_\{m\}^\{q\}\\,\\\|\\,M^\{q\}\\right\)\.\(39\)
The reported transition\-distribution JS divergence is the mean across eligible common questions:

TransitionJS=1\|𝒬common\|​∑q∈𝒬commonJStransq\.\\operatorname\{TransitionJS\}=\\frac\{1\}\{\|\\mathcal\{Q\}\_\{\\mathrm\{common\}\}\|\}\\sum\_\{q\\in\\mathcal\{Q\}\_\{\\mathrm\{common\}\}\}\\operatorname\{JS\}\_\{\\mathrm\{trans\}\}^\{q\}\.\(40\)
Lower values indicate that the model more closely reproduces the human population\-level distribution of response persistence and change across adjacent waves\.

## Appendix HAdditional Results

Table 10:Robustness of the LifeMem ablations across respondent samples\. Results report means across three independently sampled respondent sets generated with seeds 42, 43, and 44, with the corresponding±\\pmstandard deviations shown as subscripts\. KL Div\., WG Gap, and Ent\. Gap denote KL divergence, within\-group pairwise distance gap, and normalized entropy gap\. All metrics are lower\-is\-better\. Bold marks the best mean result for each model and metric; LifeMem \(Full\) is shaded light gray\.Table 11:Valid response rates across different methods and backbone models on Add Health and Understanding Society\. A response is considered valid when it can be mapped to one of the predefined answer options\. Higher values indicate fewer missing or unparsable responses\.### Robustness Across Respondent Samples

To evaluate robustness to respondent sampling, we repeat the main experiments using three independently sampled respondent sets generated with seeds 42, 43, and 44\. Each run contains 100 respondents per dataset, while all other experimental settings remain unchanged\. Table[9](https://arxiv.org/html/2608.19621#A7.T9)reports the main\-method comparison, while Table[10](https://arxiv.org/html/2608.19621#A8.T10)reports the LifeMem ablation results\. Both tables present the mean and standard deviation across the three respondent samples\.

Table[9](https://arxiv.org/html/2608.19621#A7.T9)shows that LifeMem remains the strongest method across respondent samples, achieving the best mean result in 16 of the 18 model–dataset–metric settings\. The generally small standard deviations indicate that the main conclusions are not driven by a particular respondent sample\.

Table[10](https://arxiv.org/html/2608.19621#A8.T10)further shows that the full LifeMem model consistently outperforms both ablations\. This result supports the complementary contributions of structured event memory and persistent parametric memory across different respondent samples\.

### Valid Response Rates

Table[11](https://arxiv.org/html/2608.19621#A8.T11)shows that all LLM\-based methods achieve high valid response rates across models and datasets, generally exceeding those in the human survey data\. The lower human rates reflect refusals, nonresponse, and other missing\-value cases that naturally occur in real\-world surveys\. The consistently high rates across methods also indicate that the experimental framework reliably produces responses that can be mapped to the predefined answer options\.

## Appendix IMechanism and Representation Analyses

### LoRA Visualization Procedure

To visualize the evolution of agent\-specific parametric memory, we extract the LoRA parameters of every agent after each survey wave\. For agentiiat wavett, all LoRA parameter tensors are flattened and concatenated into a single vector:

𝜽i,t=concatm∈ℳ⁡vec⁡\(𝚯i,t\(m\)\),\\boldsymbol\{\\theta\}\_\{i,t\}=\\operatorname\{concat\}\_\{m\\in\\mathcal\{M\}\}\\operatorname\{vec\}\\\!\\left\(\\boldsymbol\{\\Theta\}\_\{i,t\}^\{\(m\)\}\\right\),\(41\)whereℳ\\mathcal\{M\}denotes the consistently ordered set of LoRA tensors, including the correspondinglora\_Aandlora\_Bparameters across adapted modules\. Using a fixed ordering ensures that each dimension has the same parameter identity across agents and waves\.

We stack the vectors from allNNagents andTTwaves into a single matrix:

𝐗=\[𝜽1,1⊤𝜽1,2⊤𝜽N,T⊤\]\.\\mathbf\{X\}=\\begin\{bmatrix\}\\boldsymbol\{\\theta\}\_\{1,1\}^\{\\top\}\\\\ \\boldsymbol\{\\theta\}\_\{1,2\}^\{\\top\}\\\\ \\vdots\\\\ \\boldsymbol\{\\theta\}\_\{N,T\}^\{\\top\}\\end\{bmatrix\}\.\(42\)Each parameter dimension is then standardized globally across all agent–wave states:

Xj,d′=Xj,d−μdσd\+ϵ,X^\{\\prime\}\_\{j,d\}=\\frac\{X\_\{j,d\}\-\\mu\_\{d\}\}\{\\sigma\_\{d\}\+\\epsilon\},\(43\)whereμd\\mu\_\{d\}andσd\\sigma\_\{d\}are the mean and standard deviation of dimensionddover all rows of𝐗\\mathbf\{X\}, andϵ\\epsilonis a small constant for numerical stability\. This shared normalization prevents dimensions with larger numerical scales from dominating the projection\.

We fit a single two\-dimensional PCA projection to the globally standardized matrix:

𝐙=PCA2⁡\(𝐗′\)\.\\mathbf\{Z\}=\\operatorname\{PCA\}\_\{2\}\(\\mathbf\{X\}^\{\\prime\}\)\.\(44\)The resulting point𝐳i,t∈ℝ2\\mathbf\{z\}\_\{i,t\}\\in\\mathbb\{R\}^\{2\}represents the LoRA state of agentiiafter wavett\. For each agent, states are connected in chronological order,

𝐳i,1→𝐳i,2→⋯→𝐳i,T,\\mathbf\{z\}\_\{i,1\}\\rightarrow\\mathbf\{z\}\_\{i,2\}\\rightarrow\\cdots\\rightarrow\\mathbf\{z\}\_\{i,T\},\(45\)forming a longitudinal parameter trajectory\. Point colors indicate survey waves, while connecting lines identify states belonging to the same agent\.

### Additional LoRA Visualizations

Table 12:Effect of the temporal decay coefficientαret\\alpha\_\{\\mathrm\{ret\}\}in LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. Lower values are better for all metrics\. Bold marks the best result for each metric, and the setting used in the main experiments is shaded light gray\.Table 13:Effect of the replay loss weightη\\etain LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. Lower values are better for all metrics\. Bold marks the best result for each metric, and the setting used in the main experiments is shaded light gray\.Figures[8](https://arxiv.org/html/2608.19621#A10.F8)–[12](https://arxiv.org/html/2608.19621#A10.F12)extend the*Evolution of Agent\-Specific LoRA States*analysis in the main paper, which presents only the Understanding Society results with Llama\-3\.1\-8B\-Instruct\. Here, we provide the corresponding visualizations for all remaining model–dataset combinations: the three backbone models on Add Health and Ministral\-3\-8B\-Instruct\-2512 and Qwen3\.5\-9B on Understanding Society\. Each point represents an agent’s LoRA state after a survey wave, while connected points trace the same agent over time\. Together, these figures show whether the longitudinal evolution of agent\-specific parametric states observed in the main\-paper example also appears across other backbone models and datasets\.

### LoRA States across Life\-Event Coverage

Figure[7](https://arxiv.org/html/2608.19621#A10.F7)compares the LoRA trajectories obtained when each wave is allowed to use at most 5, 20, 40, or 90 life events\.

With only a small number of life events, the adapter trajectories remain comparatively concentrated\. As event availability increases, later\-wave states show clearer agent\-specific differentiation\.

## Appendix JHyperparameter and Design Analyses

### Temporal Decay

We examine the temporal decay coefficientαret\\alpha\_\{\\mathrm\{ret\}\}in LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. We varyαret\\alpha\_\{\\mathrm\{ret\}\}over\{0\.5,0\.7,0\.8,0\.9,0\.95,1\.0\}\\\{0\.5,0\.7,0\.8,0\.9,0\.95,1\.0\\\}while keeping all other settings fixed\. A larger value assigns relatively more weight to earlier life events, whereas a smaller value emphasizes more recent events\.

Table[12](https://arxiv.org/html/2608.19621#A9.T12)suggests that an intermediate temporal decay coefficient provides the most favorable overall performance among the evaluated settings\. In particular,αret=0\.9\\alpha\_\{\\mathrm\{ret\}\}=0\.9achieves the lowest KL divergence and entropy gap\. Smaller values may discount earlier events too strongly, whereas values closer to 1 preserve older events with little decay\. These results suggest that balancing recent and earlier experiences is useful for maintaining both distributional and diversity alignment\.

### Replay

Table 14:Effect of retriever choice onEvent RAGand LifeMem with Ministral\-3\-8B\-Instruct\-2512\. Lower values are better for all metrics\. Bold marks the best overall result, and LifeMem is shaded light gray\.Table 15:Effect of the LoRA rankrrin LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. Lower values are better for all metrics\. Bold marks the best result for each metric, and the setting used in the main experiments is shaded light gray\.Figure 7:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Add Health withLlama\-3\.1\-8B\-Instructunder maximum life\-event counts of 5, 20, 40, and 90\. Points denote LifeMem agent states after each survey wave, and connected trajectories trace the same agent over time\. With fewer events, different agents remain relatively concentrated; richer histories produce clearer agent\-specific differentiation in later waves\.We examine the replay loss weightη\\etain LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. We varyη\\etaover\{0,0\.25,0\.5,0\.75,1\.0\}\\\{0,0\.25,0\.5,0\.75,1\.0\\\}while keeping all other settings fixed\. A largerη\\etaplaces greater emphasis on replayed historical events during optimization, whereas a smallerη\\etaprioritizes examples from the current wave\. The settingη=0\\eta=0disables the contribution of replay examples to the training loss\.

Table[13](https://arxiv.org/html/2608.19621#A9.T13)suggests that an intermediate replay weight provides a favorable trade\-off across the evaluated metrics\. In particular,η=0\.5\\eta=0\.5yields the lowest KL divergence and entropy gap, although smaller replay weights achieve a lower WG Gap\. This pattern indicates that replay can support the retention of earlier experiences, but assigning it excessive or insufficient weight does not consistently improve all evaluation dimensions\.

### LoRA Rank

We examine the LoRA rankrrin LifeMem on Add Health with Llama\-3\.1\-8B\-Instruct\. We varyrrover\{4,8,16,32\}\\\{4,8,16,32\\\}while keeping all other settings fixed\. A larger rank increases the capacity of the agent\-specific adapter, but also raises its training and storage cost\.

Table[15](https://arxiv.org/html/2608.19621#A10.T15)shows that performance generally improves as the LoRA rank increases, withr=32r=32achieving the lowest values across all three metrics\. The improvement is especially clear for KL divergence and WG Gap, suggesting that additional adapter capacity may better support the representation of heterogeneous longitudinal information\. We nevertheless user=8r=8in the main experiments because it provides a more practical balance between predictive performance, computational cost, and per\-agent storage, which becomes important when simulating many agents across multiple waves\.

### Retriever Choice

We compare several retrieval backends forEvent RAGon Add Health and Understanding Society with Ministral\-3\-8B\-Instruct\-2512\. The evaluated retrievers include BM25,all\-MiniLM\-L6\-v2,E5\-large\-v2,BGE\-M3, andBGE\-M3followed byBGE\-Reranker\-v2\-M3\. We additionally report LifeMem, which usesall\-MiniLM\-L6\-v2for structured event retrieval while maintaining an agent\-specific parametric memory\.

Table[14](https://arxiv.org/html/2608.19621#A10.T14)shows that retriever choice affectsEvent RAG, but no single retriever performs best on both datasets\.all\-MiniLM\-L6\-v2performs best among theEvent RAGvariants on Add Health, whileBGE\-M3with reranking performs best on Understanding Society\. LifeMem achieves lower values across all metrics while using the sameall\-MiniLM\-L6\-v2retriever, suggesting that its gains are not solely due to retriever choice\.

![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_ah_llama.png)Figure 8:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Add Health with Llama\-3\.1\-8B\-Instruct\. Points represent agent states after each survey wave, and connected trajectories trace the same agent over time\. Colors indicate survey waves\.![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_ah_mistral.png)Figure 9:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Add Health with Ministral\-3\-8B\-Instruct\-2512\. Points represent agent states after each survey wave, and connected trajectories trace the same agent over time\. Colors indicate survey waves\.![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_ah_qwen.png)Figure 10:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Add Health with Qwen3\.5\-9B\. Points represent agent states after each survey wave, and connected trajectories trace the same agent over time\. Colors indicate survey waves\.![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_us_mistral.png)Figure 11:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Understanding Society with Ministral\-3\-8B\-Instruct\-2512\. Points represent agent states after each survey wave, and connected trajectories trace the same agent over time\. Colors indicate survey waves\.![Refer to caption](https://arxiv.org/html/2608.19621v1/lora_us_qwen.png)Figure 12:PCA visualization of agent\-specific LoRA states learned by LifeMem for 100 agents on Understanding Society with Qwen3\.5\-9B\. Points represent agent states after each survey wave, and connected trajectories trace the same agent over time\. Colors indicate survey waves\.