PaperlayerView exact revision on arXiv

Source-linked AI summary

AnnaAgent: Dynamic Evolution Agent System with Multi-Session Memory for Realistic Seeker Simulation

Ming Wang, Peidong Wang, Lin Wu, Xiaocui Yang, Daling Wang, Shi Feng, Yuxin Chen, Bixuan Wang, Yifei Zhang

arXiv:2506.00551v2cs.CLcs.AI

TL;DR

Seeker simulation faces challenges in modeling dynamic evolution and multi-session memory. AnnaAgent addresses these with emotion and complaint control plus tertiary memory, and evaluations indicate more realistic simulations than baselines.

  • Problem

    Realistic seeker simulation remains challenged by modeling dynamic evolution and multi-session memory.

    Source: Limitation · p. 9

  • Method

    AnnaAgent uses an emotion modulator and chief complaint elicitor for within-session evolution, alongside tertiary memory for multi-session memory.

    Source: §1 · p. 2, §3.1 · p. 4, Limitation · p. 9

  • Results

    AnnaAgent experimentally simulates seekers more realistically than baseline methods.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The authors present AnnaAgent as an approach that provides innovative solutions to alleviating global mental-health resource shortages.

    Source: §5 · p. 9

    The dynamic-evolution practices remain formally simplified, and coordination of multi-session memory through tertiary memory remains rudimentary.

    Source: Limitation · p. 9

Abstract

from arXiv · show

Constrained by the cost and ethical concerns of involving real seekers in AI-driven mental health, researchers develop LLM-based conversational agents (CAs) with tailored configurations, such as profiles, symptoms, and scenarios, to simulate seekers. While these efforts advance AI in mental health, achieving more realistic seeker simulation remains hindered by two key challenges: dynamic evolution and multi-session memory. Seekers' mental states often fluctuate during counseling, which typically spans multiple sessions. To address this, we propose AnnaAgent, an emotional and cognitive dynamic agent system equipped with tertiary memory. AnnaAgent incorporates an emotion modulator and a complaint elicitor trained on real counseling dialogues, enabling dynamic control of the simulator's configurations. Additionally, its tertiary memory mechanism effectively integrates short-term and long-term memory across sessions. Evaluation results, both automated and manual, demonstrate that AnnaAgent achieves more realistic seeker simulation in psychological counseling compared to existing baselines. The ethically reviewed and screened code can be found on https://github.com/sci-m-wang/AnnaAgent.

1 Introduction

Existing seeker simulators struggle to capture emotional and complaint changes during counseling and to retain information across sessions. AnnaAgent combines dynamic control with staged memory and is reported to simulate seekers more realistically than baselines.

1 Introduction

  • Real seekers are costly and raise ethical concerns in research, motivating LLM-based conversational agents configured with seeker profiles and symptoms.Such agents are used to simulate conversational behavior of real seekers.Source: §1 · p. 1
  • Existing simulators may hold seekers’ emotions at their initial state and lack multi-session memory, risking confusion or hallucinations about prior-session topics.Counselors use earlier-session topics to review suggestions, discuss changes, build trust, and assess progress.Source: §1 · pp. 1–2
  • AnnaAgent is reported to simulate seekers more realistically than baselines on anthropomorphism, personality fidelity, and previous-session cognitive accuracy.The introduction names these three evaluation dimensions.Source: §1 · p. 2
  • The paper frames dynamic evolution as changes in emotions and complaints, and multi-session memory as staged memory; AnnaAgent controls those changes and schedules memories.Its system is described as controlling emotional and symptomatic cognitive changes while using tertiary memory.Source: §1 · p. 2

2 Related Works

Prior work spans mental-health dialogue, traditional seeker simulation, and LLM role-playing. The related-work discussion contrasts the cost and realism of human role-play with the potential of coordinated agents to address static configurations.

2 Related Works

  • LLMs are used in mental-health dialogue systems, while this work focuses on seeker simulation and also relates to role-playing and multi-agent systems.The cited applications include counseling and support dialogue systems.Source: §2 · p. 3

2.1 Traditional Seeker Simulation

  • Rogers et al. (2022) found standardized patients better simulate actual counseling than virtual seekers, though at much higher cost.The discussion also describes standardized patients and peer role-playing as useful in communication training.Source: §2.1 · p. 3

2.2 LLM-based Role-Playing and Multi-Agent System

  • Role-profile datasets and personality-conditioned tuning strengthen role-play, but static single-agent configurations remain constrained by the two challenges, motivating AnnaAgent’s collaborative multi-agent design.The proposed collaboration is intended to control dynamic evolution and schedule multi-session memories.Source: §2.2 · p. 3

3 Challenge Formalization and AnnaAgent Design

AnnaAgent assigns conversational-agent configurations to two agent groups: one controls changing seeker emotions and complaints, and the other schedules memories across sessions. Its emotion and complaint mechanisms model change over a session, while tertiary memory separates and retrieves information by timing and session.

3.1 Designing Conversational Agents for Seeker Simulation

  • AnnaAgent keeps profiles static, reserves slots for changing emotions, complaints, status, and situation, and assigns two agent groups to evolution control and multi-session memory scheduling.The configuration table distinguishes static profile information from fields controlled by dynamic evolution or scheduled memory.Source: §3.1 · pp. 3–4, Table 1 · p. 4

3.2 Dynamic Evolution

  • A reminder after each dialogue round supplies the virtual seeker’s current emotion and chief complaint state.Source: §3.2 · p. 4

3.2.1 Emotion Modulation

  • The emotion model predicts the next seeker utterance’s emotion from the seeker’s profile and existing conversation, using a model trained on D4 counseling dialogues.Training instances use GPT-4o emotion labels and omit the final seeker utterance so the model learns to predict it.Source: §3.2.1 · p. 5
  • To avoid fixed emotional patterns, AnnaAgent randomly perturbs inferred emotions, assigning higher probability to emotions in closer groups; volatility is usually minor without strong stimuli.Source: §3.2.1 · p. 5

3.2.2 Chief Complaint Elicitor

  • Before a session, a generator uses the seeker configuration and a recent event to produce a complaint-change chain with multiple stages.Source: §3.2.2 · p. 5
  • D4-trained complaint chains receive GPT-4o annotation and psychology-expert review; during conversation, the elicitor judges whether to advance the complaint to the next stage.Chains are manually corrected when more than two annotators consider them unreasonable.Source: §3.2.2 · pp. 5–6

3.3 Tertiary Memory System

  • Tertiary memory divides information into real-time conversation, short-term recent events and status changes, and long-term prior-session experiences, scales, and conversations.Content through the last session is long-term; content before the current session is real-time; intervening content is short-term.Source: §3.3 · p. 6
  • Short-term memory uses self-report scales and profile-matched recent events, while long-term retrieval supplements a response when the querier detects a previous-session topic.The retrieved material comes from prior-session conversations and scale records; real-time memory is supplied as the seeker model’s full context.Source: §3.3 · p. 6

4 Experiments

The experiments evaluate AnnaAgent against baseline methods across seeker-simulation metrics, component ablations, and backbone variation.

4.1 Experiment Settings

  • The study uses records in D4 and DAIC-WOZ as the basis for previous sessions and uses both datasets as seeds to generate reports with GPT-4o-mini (OpenAI et al., 2024).Raw-data-related code will be conditionally available.Source: §4.1 · p. 6

4.2 Metrics

  • Anthropomorphism is the portfolio-average BERT-score against real-seeker utterances, while personality fidelity uses G-Eval on answers to profile-oriented interview questions.The BERT-score calculation randomly samples 10% of candidate-reference sentence pairs.Source: §4.2 · p. 7

4.3 Main Results

  • AnnaAgent achieves the best anthropomorphism in nearly all tested counselor and dataset conditions, and is overall superior on personality fidelity.It ranks second for conversations with PsycoLLM on D4; Figure 4 reports personality-fidelity scores.Source: §4.3 · p. 7, Figure 4 · p. 7, Table 2 · p. 8

4.5 Generalizability Study

  • Across backbone models, all RSDs were below 10%, and AnnaAgent was less affected by backbone choice, indicating reported stability and generalization.Table 3 reports performance using different backbone architectures.Source: Table 3 · p. 8, §4.5 · p. 9

4.4 Ablation Study

  • Without dynamic evolution, performance significantly degrades; ablating long-term memory significantly reduces accuracy on prior-session cognitive-perception questions.The long-term-memory ablation is evaluated using preset questions scored by G-Eval.Source: Table 4 · p. 8, §4.4 · pp. 8–9, Figure 5 · p. 8

5 Conclusion

AnnaAgent addresses dynamic evolution and multi-session memory in seeker simulation through mechanisms for controlling emotional and cognitive changes and coordinating memory. The conclusion reports more realistic simulations than baseline methods, while noting component necessity and generalization evaluations.

  • The paper frames dynamic evolution and multi-session memory as challenges to more realistic seeker simulation.Source: §5 · p. 9
  • It uses real-time emotional inference with random perturbations and complaint-elicitation chains to control seekers’ emotional and symptomatic cognitive changes.A tertiary memory system coordinates memory during and between sessions.Source: §5 · p. 9
  • AnnaAgent more realistically simulates seekers than baseline methods, with evaluations also demonstrating component necessity and generalization.The conclusion reports these outcomes without specifying quantitative results.Source: §5 · p. 9

Limitation

The paper introduces dynamic evolution and multi-session memory as issues in seeker simulation and proposes AnnaAgent to address them. It also acknowledges simplifications in its modeling and identifies further work on evolution and memory coordination.

  • AnnaAgent addresses dynamic evolution within sessions through an emotion modulator and chief complaint elicitor, and coordinates multi-session memory with a tertiary memory system.Source: Limitation · p. 9
  • The practices remain formally simplified, leaving finer-grained modeling of counseling dynamics and memory coordination as needed improvements.The paper specifically describes tertiary-memory coordination as rudimentary.Source: Limitation · p. 9
  • Future work will investigate navigating dynamic evolution during a counseling session and coordinating memory across multiple sessions.Source: Limitation · p. 9

Ethical Considerations

The ethical considerations cover privacy risks from real counseling data, potential misuse of seeker simulations, and the scope of released data. The section also points to baseline prompt templates and a simulator template whose slots are filled with AnnaAgent-controlled configurations.

Ethical Considerations

  • Training components and prior-session inputs use real counseling data containing user profiles, creating a possible patient-privacy leakage risk despite omission of key information.The authors report signing a dataset application agreement.Source: Ethical Considerations · p. 9
  • AnnaAgent can be used to generate counseling conversations, train counselor models, or simulate experiments, but it is not a complete substitute for real seekers.The authors position it as a supplementary tool and report an ethical review by experienced counselors.Source: Ethical Considerations · p. 10

A Baseline Prompts

B Prompt Template for Simulators

C Algorithm of the Chief Complaint Elicitor

The section identifies Algorithm 1 as the procedure for deciding whether to switch chief-complaint stages. Its named functions generate a complaint-change chain and assess whether the seeker has recognized the current complaint stage.

D Questions for Personality Fidelity

The section describes a seven-question evaluation of personality fidelity in seeker simulation methods, including a question about core concerns affecting mood and sleep.

D Questions for Personality Fidelity

2 Can you describe your experience from wak-

The questions probe interpersonal patterns, beliefs, strengths, problem origins, and hoped-for changes, covering several areas relevant to understanding a person's personality.

2 Can you describe your experience from wak-

  • One question asks when the person first clearly realized a problem existed.The question set describes tracing a problem's origins as one area of exploration.Source: §2 · p. 14
  • Another asks what feelings or behaviors would change first if counseling succeeded.The set frames this as envisioning minimal future changes.Source: §2 · p. 14
  • The questions cover interpersonal interactions, automatic thoughts when misunderstood, and strengths recognized by family or friends.These topics probe interpersonal patterns, core beliefs, and external perspectives on strengths.Source: §2 · p. 14
  • Together, the questions span core disturbances, specific life slices, interpersonal interactions, beliefs, strengths, problem origins, and future changes.Source: §2 · p. 14

E Questions for Long-term Memory

The section describes seven questions for evaluating long-term memory and explains that explicitly prompting recall of prior-session experiences is intended to increase memory activation.

E Questions for Long-term Memory

Loading 2506.00551v2…