A new preprint reports that language-model agents given a running record of life events matched human survey response patterns more closely than agents built from static profiles or simpler memory strategies. In the tested settings, LifeMem also showed smaller gaps in response diversity and, on the common Understanding Society longitudinal set, a closer match for how answers changed between waves.
In an auxiliary WVS diagnostic, 2,000 randomly sampled respondents were divided into three socioeconomic groups. A silhouette score - a measure of how tightly items cluster within groups and how distinct those groups are - was minus 0.02 for human responses and 0.19 for profile-conditioned agents. The pattern pointed to more overlap among human groups but greater within-group compression and between-group separation among agents.
A memory built for change
LifeMem pairs structured event memory, preserving event content, timing and evidence for retrieval, with agent-specific LoRA adapters - a targeted way of updating part of a model. Those adapters are updated separately for each agent so experience can persist across questions and waves while the base model stays frozen.
The main evaluation converted public-use, de-identified longitudinal records from Add Health and Understanding Society into respondent-derived model agents. It used six Add Health waves and 15 Understanding Society waves, kept only complete all-wave cases, and sampled 100 respondents per dataset from aligned pools of 2,048 and 14,104.
The comparison included baselines using direct instructions, static profiles, full histories and retrieved events. Researchers used four lower-is-better measures: the distance between human and model answer distributions, the gap in within-group spread, the gap in overall response diversity, and the divergence in how answers changed between adjacent waves. Lower values meant closer alignment.
Better averages, with trade-offs
LifeMem generally showed smaller distributional and diversity gaps than the alternatives. In the reported LifeMem row, KL divergence - the difference between answer distributions - was 4.0635 for Add Health and 3.4529 for Understanding Society. The within-group gaps were 0.2309 and 0.2886, the response-diversity entropy gaps were 0.3207 and 0.3742, and the Understanding Society transition JS divergence was 0.3331. The paper reports largely significant gains; significance annotations used two-sided paired t-tests with p < 0.05.
The pattern was not uniform. For Qwen3.5-9B on Add Health, Multilingual produced slightly smaller within-group and entropy gaps than LifeMem, although its KL divergence was higher, indicating worse overall distributional agreement.
When either structured event memory or parametric memory was removed, performance was worse on most measures, supporting the view that the two components contribute in complementary ways. Across respondent samples generated with seeds 42, 43 and 44, LifeMem had the best mean result in 16 of 18 model-dataset-metric settings and consistently outperformed both reduced versions.
That performance came with a cost in speed. LifeMem was slower than lightweight prompt-only baselines but faster than Full History, Event RAG and SimVBG. Its reported latency was 161.7 milliseconds plus a 0.9-millisecond adapter-loading component on Add Health, and 154.5 milliseconds plus 1.4 milliseconds on Understanding Society. A separate retrieval test found that metrics largely plateaued around 40 retrieved events, while runtime and input-token use continued to increase.
What the scores cannot establish
These are aggregate simulation results, not evidence that LifeMem creates a digital replica of any individual. The study also does not establish that identity essentialism is the cause of the clustering pattern, or that the evaluated agents are suitable for consequential individual decisions. Whether aggregate distributional and diversity measures predict individual-level realism remains an open question.
The main experiments used only 100 respondent-derived agents per dataset, even though the complete-case aligned pools were larger. Keeping only people with data for every selected wave can exclude respondents who missed a wave. Metrics also dropped missing, refusal, do-not-know and not-applicable human responses, so the comparisons were conditional on valid human-model pairs.
Paper data and sources
Original title: Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories
Authors: Hexi Wang, Yujia Zhou, Bangde Du et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text