AnnaAgent

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

Venue: ACL 2025 (Findings)
Paper: ACL Anthology

Overview

AnnaAgent aims to simulate the dynamic evolution of the AI seeker's emotions and multi-session memory. It features a sophisticated emotion system based on the GoEmotions taxonomy (28 emotion categories) and models the progression of psychological complaints through multi-stage chains.

Key Features

• Multi-Session Memory: Maintains context across multiple therapy sessions
• Dynamic Evolution: Complaint progression through cognitive change chains
• 28-Emotion System: Based on GoEmotions taxonomy (Positive, Neutral, Ambiguous, Negative)
• Profile-Based: Detailed patient profiles with depression/suicide risk levels
• Previous Session Integration: References historical conversations when relevant

How It Works

1. Profile Loading: Loads detailed patient profile with risk levels and symptoms
2. Emotion Inference: Determines emotional state based on therapist input
3. Context Check: Evaluates if historical information is needed for the response
4. Complaint Progression: Tracks progress through the cognitive change chain
5. Response Generation: Produces contextually appropriate responses with emotional consistency

Emotion Categories

AnnaAgent uses the GoEmotions taxonomy with weighted emotion transitions:

Category	Emotions
Positive	admiration, amusement, approval, caring, curiosity, excitement, gratitude, joy, love, optimism, pride, relief
Neutral	neutral
Ambiguous	confusion, disappointment, nervousness
Negative	anger, annoyance, disapproval, disgust, embarrassment, fear, sadness, remorse, grief

Usage

CLI

patienthub simulate client=annaAgent

Python

from patienthub.clients import get_client

client = get_client(agent_name="annaAgent", lang='en')

response = client.generate_response("How have you been since our last session?")
print(response)

The original implementation described in this paper are based on Qwen2.5-7B-Instruct.

Configuration

Option	Description	Default
prompt_path	Path to prompt file	data/prompts/client/annaAgent.yaml
data_path	Path to character file	data/characters/annaAgent.json
data_idx	Character index	0

Character Data Format

{
  "profile": {
    "age": "42",
    "gender": "Female",
    "marital_status": "Divorced",
    "occupation": "Teacher",
    "symptoms": "Lack of self-confidence, low self-worth, feelings of guilt..."
  },
  "situation": "Description of the current situation triggering distress...",
  "statement": ["Initial statements the patient might make..."],
  "style": ["Self-critical", "Emotionally restrained", "Direct and concise"],
  "complaint_chain": [
    { "stage": 1, "content": "Initial complaint..." },
    { "stage": 2, "content": "Deeper realization..." },
    { "stage": 3, "content": "Core belief surfaces..." }
  ],
  "status": "Summary of patient's current mental health status...",
  "report": {
    "case_title": "...",
    "case_categories": ["Anxiety Disorder", "Low Self-Worth"],
    "techniques_used": ["Cognitive Behavioral Therapy", "Emotional Support"],
    "case_summary": ["..."],
    "counseling_process": ["..."],
    "insights_and_reflections": ["..."]
  },
  "previous_conversations": [
    { "role": "Client", "content": "..." },
    { "role": "Therapist", "content": "..." }
  ]
}

Resources

data/resources/annaAgent_events.json: Psychological distress triggering events, keyed by language (en/zh). Each has adult (age-tagged {age, event} records, used for ±5-year age matching) and teen (a list of event strings).

GHQ, SASS, and BDI scales are shared constants in patienthub.resources (ANNAAGENT_SCALES).