anna-agent 0.2.2

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

AnnaAgent

中文说明 | English

The code for the paper AnnaAgent: Dynamic Evolution Agent System with Multi-Session Memory for Realistic Seeker Simulation.

Important repository update: We have removed the previous web application code and refactored this repository into a cleaner AnnaAgent core toolkit so that the method is easier to understand, install, and use for experiments or downstream applications. If you need the web application, please use the deployed site at anna.kinamind.org or the AnnApod repository at github.com/kinamind/annapod.

CLAIM

It is important to note that since this work involves data from counselling records of real patients with psychological disorders, the publicly available code can only be used to demonstrate part of the methodology. Please contact the authors of this paper if needed.

Installation

Install as a terminal command

For most readers, install AnnaAgent from PyPI as a terminal command with uv or pipx:

# Install the latest PyPI release as a standalone tool.
uv tool install anna-agent

# Or, if you use pipx:
pipx install anna-agent

# Or, if you prefer pip in an existing environment:
pip install anna-agent

After installation, the short command anna is available in any terminal:

anna --version
anna create anna-workspace
anna doctor --workspace anna-workspace

The longer command name anna-agent is kept as a compatibility alias.

If you want AnnaAgent to automatically start local SFT models with vLLM, create a workspace deployment environment on the Linux/GPU machine. This keeps the lightweight anna CLI install separate from heavy vLLM dependencies:

anna create anna-workspace --deploy-env

# Or add it later to an existing workspace.
anna models env setup --workspace anna-workspace
anna models env status --workspace anna-workspace

The deploy environment is created under anna-workspace/.anna-deploy-venv with Python 3.12 by default. anna models deploy automatically uses anna-workspace/.anna-deploy-venv/bin/vllm when it exists. If your cluster already provides vLLM in another environment, keep the normal AnnaAgent install and pass that executable explicitly when deploying:

anna models deploy --target complaint --workspace anna-workspace \
  --vllm-command /path/to/vllm

Develop from source

If you are modifying the code, install the project dependencies into a project-local .venv using uv:

git clone https://github.com/sci-m-wang/AnnaAgent.git
cd AnnaAgent
uv sync
uv run anna --help

The full bilingual documentation is available from the GitHub Pages site: https://sci-m-wang.github.io/AnnaAgent/.

You can also expose the local checkout as the terminal command:

uv tool install --editable .
anna --help

How to Run the Example

First, create a workspace, choose whether to use the SFT modules, and let the CLI write the resulting configuration:

anna create anna-workspace

# Optional on GPU machines: create anna-workspace/.anna-deploy-venv for vLLM.
anna create anna-workspace --deploy-env

# Fast path: use the base chat model for complaint-chain and emotion modules.
anna models use-base --target all --workspace anna-workspace

# SFT path: deploy local vLLM services from downloaded HuggingFace assets.
anna assets download paper --workspace anna-workspace
anna models env setup --workspace anna-workspace  # optional if create used --deploy-env
anna models deploy --target complaint --backend vllm --workspace anna-workspace \
  --gpu 0 --gpu-memory-utilization 0.85 --wait-timeout 900
anna models deploy --target emotion --backend vllm --workspace anna-workspace \
  --gpu 1 --gpu-memory-utilization 0.85 --wait-timeout 900

You can also connect self-hosted OpenAI-compatible SFT endpoints instead of letting AnnaAgent start vLLM:

anna models configure --target complaint \
  --base-url http://127.0.0.1:8001/v1 \
  --model-name complaint \
  --workspace anna-workspace

anna models configure --target emotion \
  --base-url http://127.0.0.1:8000/v1 \
  --model-name emotion \
  --workspace anna-workspace

Using the base model is easier to start with. The SFT modules usually produce more faithful emotion inference and chief-complaint chains when GPU resources are available.

There is an inner example provided through the anna CLI. Install the dependencies and initialize the project before starting the demo:

uv run python -m anna_agent.initialize
uv run anna

After initialization you can chat with the virtual seeker.

CLI Usage

AnnaAgent provides a Typer-based CLI organized around the reader journey from paper reproduction to application use. Start by creating an isolated workspace:

anna create anna-workspace
anna doctor --workspace anna-workspace

The workspace contains settings.yaml, .env, sample cases, prompts, run outputs, logs, cache files and an asset manifest. Configure model endpoints with the wizard or non-interactive setters:

anna config wizard --workspace anna-workspace
anna config secrets --workspace anna-workspace
anna config set model_service.base_url https://example.com/v1 \
  --workspace anna-workspace
anna config show --workspace anna-workspace
anna config validate --workspace anna-workspace

config wizard and config secrets use hidden password-style prompts for API keys and write them to .env. The generated .env and .env.example files include commented placeholders showing exactly where to put backbone, SFT and embedding credentials if you prefer manual editing.

Assets are manifest-driven. The default paper preset points to the released HuggingFace SFT models and synthetic data, and you can override or extend it in assets/anna-assets.json with your own repositories or direct URLs:

anna assets list --workspace anna-workspace
anna assets download paper --workspace anna-workspace

Always pass --workspace or --manifest when downloading assets. Without either, anna assets download uses the current directory as the workspace and may download to ./assets/... instead of your intended AnnaAgent workspace. You can download one specific resource or override the target directory explicitly:

# Download one asset from anna-workspace/assets/anna-assets.json.
anna assets download complaint-sft --workspace anna-workspace

# Use an explicit manifest JSON, including absolute target paths in that file.
anna assets download complaint-sft --manifest anna-workspace/assets/anna-assets.json

# Override the target directory for exactly one selected asset.
anna assets download complaint-sft --workspace anna-workspace \
  --target /path/to/models/complaint-sft

Choose the model mode explicitly before experiments. Use the base model for the lowest setup cost, configure existing SFT endpoints if you already deployed them, or let AnnaAgent start local vLLM services:

anna models use-base --target all --workspace anna-workspace
anna models use-sft --target all --workspace anna-workspace
anna models status --workspace anna-workspace

anna models configure --target complaint \
  --base-url http://127.0.0.1:8001/v1 \
  --model-name complaint \
  --workspace anna-workspace

anna models env setup --workspace anna-workspace
anna models env status --workspace anna-workspace
anna models deploy --target complaint --backend vllm --workspace anna-workspace \
  --gpu 0 --gpu-memory-utilization 0.85 --wait-timeout 900
anna models deploy --target emotion --backend vllm --workspace anna-workspace \
  --gpu 1 --gpu-memory-utilization 0.85 --wait-timeout 900

models deploy starts a vLLM OpenAI-compatible server in the background, writes the service URL/model name/use-SFT flag back to settings.yaml, writes API keys to .env, and records logs/PIDs under logs/services/ and runs/services/. Before starting vLLM, it runs a GPU preflight check with nvidia-smi, prints the selected GPU, free memory, and vLLM memory cap, and blocks obvious failures such as a missing GPU ID or insufficient free memory. Before writing configuration, it waits for the service to answer /v1/models; use --wait-timeout 900 for slow model loads. If startup fails or times out, the CLI prints the service log tail and does not write a bad endpoint. Use --dry-run to print the vLLM command without starting anything. It also checks for a CUDA toolkit (nvcc) without assuming a fixed CUDA module name or version. If a valid toolkit is found through --cuda-home, CUDA_HOME, PATH, or common CUDA roots, AnnaAgent injects CUDA_HOME, PATH, and LD_LIBRARY_PATH only into the vLLM child process. If no toolkit is visible on a module-based cluster, AnnaAgent inspects available CUDA modules and auto-loads the default CUDA module for the vLLM process; if no default is marked, it uses the highest available version. If no toolkit or CUDA module can be found, the CLI warns and continues because some vLLM environments do not require nvcc; FlashInfer JIT environments usually do. No manual module load is required. The deploy preflight also checks for the ninja build tool required by FlashInfer JIT. If the workspace deploy environment is missing it, models deploy installs it into that workspace before starting vLLM. Use --cuda-home only when the cluster stores CUDA in a custom location:

anna models deploy --target complaint --backend vllm --workspace anna-workspace \
  --gpu 0 --gpu-memory-utilization 0.85 --wait-timeout 900

anna models deploy --target complaint --backend vllm --workspace anna-workspace \
  --gpu 0 --cuda-home /path/to/cuda --gpu-memory-utilization 0.85

When --model-path is omitted, deploy reads the corresponding SFT asset target from assets/anna-assets.json, including absolute paths. Pass the same --workspace or --manifest that you used during assets download.

If models deploy reports that vLLM is unavailable, run anna models env setup --workspace anna-workspace to create the workspace deploy environment, or pass --vllm-command to a vLLM executable provided by your cluster/conda environment.

Validate and prepare case data before running experiments:

anna data validate anna-workspace/cases/family_stress_case.json
anna data inspect anna-workspace/cases/family_stress_case.json
anna data sample --out anna-workspace/cases/sample.json

Run connectivity checks separately from expensive experiments:

anna test embedding --workspace anna-workspace
anna test memory --workspace anna-workspace
anna test model --workspace anna-workspace

Initialization can be run as a full AnnaAgent initialization, or as a prompt-only state for cheap dry-runs and reproducible prompt freezing:

anna init prompt-only anna-workspace/cases/family_stress_case.json \
  --out anna-workspace/prompts/family.prompt.json
anna init full anna-workspace/cases/family_stress_case.json \
  --out anna-workspace/prompts/family.full.json --workspace anna-workspace
anna init from-prompt anna-workspace/prompts/family.prompt.json

Chat interactively from either a case file or a frozen prompt state:

anna chat --workspace anna-workspace \
  --case anna-workspace/cases/family_stress_case.json \
  --save anna-workspace/runs/manual-chat.jsonl
anna chat --workspace anna-workspace \
  --state anna-workspace/prompts/family.prompt.json

Batch experiments support dry-run initialization by default and live scripted conversation when --live is supplied:

anna run batch --workspace anna-workspace \
  --case 'cases/*.json' --out anna-workspace/runs/batch
anna run batch --workspace anna-workspace \
  --case 'cases/*.json' --script scripts/counselor_messages.json \
  --live --out anna-workspace/runs/live-batch

Start the lightweight JSON API service for external experiment drivers:

anna serve --workspace anna-workspace --host 127.0.0.1 --port 8000

Diagnostics and cleanup commands are available for local workflows:

anna logs tail anna-workspace/logs/anna-agent.log
anna cache list --workspace anna-workspace
anna cache clean --workspace anna-workspace --yes
anna reset workspace --workspace anna-workspace --yes

Running anna without a subcommand still starts interactive chat from the workspace interactive.yaml; anna demo creates a sample case if needed and starts an example chat.

Project Initialization

The repository offers a small helper to generate default configuration files. Run the initialization script once before starting the example. It creates a settings.yaml, an interactive.yaml and .env in the target directory:

uv run python -m anna_agent.initialize

The generated settings.yaml contains the model service settings and per-module server configuration including API keys, base URLs and model names for the complaint, counselor and emotion modules. interactive.yaml holds a sample portrait, report and conversation history used by the main CLI. Environment variables are written to .env with the ANNA_ENGINE_ prefix for easy override.

Complete Run with a Sample Case

The repository includes a family-stress sample case at docs/family_stress_case.json:

• id: 42289a5f-bbdc-43f9-826a-9569bbbd5feb
• conversation: previous-session conversation history used as long-term memory
• report: structured counseling case report
• portrait: seeker profile and symptoms

Run a complete one-turn example with the sample case:

uv sync
uv run python -m anna_agent.initialize
rm -f interactive.yaml
cp docs/family_stress_case.json interactive.json
printf "最近一次感到伤心或者失望的时候，是什么原因导致的？\nexit\n" | \
  ANNA_ENGINE_COMPLAINT_USE_SFT_MODEL=false \
  ANNA_ENGINE_EMOTION_USE_SFT_MODEL=false \
  uv run anna

The two ANNA_ENGINE_*_USE_SFT_MODEL=false flags make the emotional inferencer and chief complaint chain generator use the base model configured in model_service. This is useful when the SFT checkpoints are unavailable.

Long-Term Memory with LanceDB

AnnaAgent stores long-term memory in a local LanceDB database. Previous-session conversations and reports are chunked into conversation_turn, conversation_window, session_summary, report_section, and report_summary records. Session metadata is stored alongside the vector table, so future runs can accumulate multiple sessions for the same seeker.

By default, memory data is written to .anna_memory/, which is ignored by Git. The embedding layer first tries the configured OpenAI-compatible embedding model and automatically falls back to a deterministic local hash embedding when the embedding service is unavailable.

Embedding credentials can use the AnnaAgent names or common OpenAI-style aliases in .env:

ANNA_ENGINE_EMBEDDING_API_KEY=...
ANNA_ENGINE_EMBEDDING_BASE_URL=https://your-embedding-endpoint/v1
ANNA_ENGINE_EMBEDDING_MODEL_NAME=your-embedding-model

# Also supported:
OPENAI_EMBEDDING_API_KEY=...
OPENAI_EMBEDDING_BASE_URL=https://your-embedding-endpoint/v1
OPENAI_EMBEDDING_MODEL=your-embedding-model

Index the sample case into long-term memory:

anna memory index docs/family_stress_case.json

Search a seeker's long-term memory:

anna memory search "胸闷和家庭压力" \
  --seeker-id 42289a5f-bbdc-43f9-826a-9569bbbd5feb

During normal interactive runs, AnnaAgent can auto-index the current interactive.json / interactive.yaml previous-session data and use retrieved memory when a counselor utterance refers to prior sessions or historical context.

Choosing Base Models or SFT Models

The emotional inferencer and chief complaint chain generator can either use the base model configured in model_service, self-hosted SFT endpoints, or local vLLM services launched by AnnaAgent. Prefer the explicit CLI commands over manual YAML edits:

anna models use-base --target all --workspace anna-workspace
anna models use-sft --target all --workspace anna-workspace
anna models configure --target emotion \
  --base-url http://127.0.0.1:8000/v1 \
  --model-name emotion \
  --workspace anna-workspace
anna models env setup --workspace anna-workspace
anna models deploy --target emotion --backend vllm --workspace anna-workspace \
  --gpu 1 --gpu-memory-utilization 0.85 --wait-timeout 900

Manual configuration is still supported. Set use_sft_model to false to use the base model, or true to call the configured SFT endpoint.

model_service:
  model_name: anna-backbone
  api_key: anna-backbone
  base_url: http://localhost:8002/v1
servers:
  complaint:
    use_sft_model: false
  emotion:
    use_sft_model: false

interactive.yaml Overview

interactive.yaml defines the virtual seeker's configuration. The main fields are:

• portrait – basic profile and psychological risk scores (e.g. drisk, srisk).
• report – case description including categories and applied techniques.
• previous_conversations – optional conversation history from earlier sessions.

A ready-to-use example can be found at docs/interactive_demo.yaml. It follows the psychological scale format used by the project and can be copied as your starting configuration. Another complete sample is docs/family_stress_case.json, which uses conversation as the previous-session conversation field and marital_status as a supported alias for martial_status.

The anna_agent package loads its configuration from the workspace directory at runtime using settings.yaml. By default the current working directory is used, but you can override the location by setting the ANNA_AGENT_WORKSPACE environment variable. When using the library programmatically you can also call anna_agent.backbone.configure(<workspace>) to load the desired configuration on demand.

Work Progress

To make it easier for readers to learn how to use it, we have added the flowchart below:

With two groups of agents (for Dynamic Evolution & Multi-session Memory, respectively), AnnaAgent simulates seekers via two main stages, i.e., the initialization stage to set the seeker's configuration (including profile, situation, symptoms, etc) and the conversation stage to interact with the counselor. The agent group for dynamic evolution contains an emotion modulator, a chief complaint chain generator and a complaint switcher. There are three agents in the agent group for multi-session memory: a situation analyzer, a status analyzer, and a memory retriever.

In addition, there are supplementary modules for speaking style analysis, scale summarization and event selection.

At the initialization stage, the seeker's basic profile and historical session conversations and reports from long-term memory are first read. The seeker's style is analyzed based on the previous session's conversations by the speaking style analysis module next. The scale summarization module summarizes historical scales based on the seeker's profile and reports. Then, the event selection module matches a suitable event based on the seeker's profile and the situation analyzer generates a situation that the seeker encounters based on the event. Meanwhile, the virtual seeker is required to complete the scales for the current session based on the current configurations and the status analyzer analyzes the change in the seeker's status based on the two groups of scales. Situations and statuses together make up short-term memory. In addition, the chief complaint chain generator generates a chief complaint chain based on the seeker's profile and long-term memory during the initialization stage.

At the conversation stage, AnnaAgent has a conversation with a counselor. For each utterance by the counselor, the memory retriever determines whether long-term memory needs to be retrieved. If it is needed, relevant information is retrieved from conversations and reports from previous sessions. In addition, the emotion modulator reasons the seeker's next emotion and performs emotion perturbation on a probability basis according to the real-time memory, i.e., the context of the conversation. After the seeker completes an utterance, the complaint switcher decides whether or not to awaken the seeker's next chief complaint stage.

Models

The training data for both the emotional inferencer and the chief complaint chain generator are derived from real data. We did not open source the labeled raw data due to ethical risk concerns. To facilitate further research and application, we set the models to be conditionally public.

Model	Link	Backbone
The Emotional Inferencer	link	Qwen2.5-7B-Instruct
Chief Complaint Chain Generator	link	Qwen2.5-7B-Instruct

In addition, we will continue to train and release emotion inferencers and chief complaint chain generators based on more models with different architectures.

Synthetic Data

We used the CPsyCounD dataset as a seed to synthesize a seeker bank that meets the requirements of the AnnaAgent format using GPT-4o-mini. It can be found at this link. We will continue to transform more data and will create more realistic seeker characters based on AnnaAgent for use in related research.

Developer Guide

For contribution guidelines refer to:

• 开发者指南 (Chinese)
• Developer Guide (English)

Citation

@inproceedings{wang-etal-2025-annaagent,
    title = "{A}nna{A}gent: Dynamic Evolution Agent System with Multi-Session Memory for Realistic Seeker Simulation",
    author = "Wang, Ming  and
      Wang, Peidong  and
      Wu, Lin  and
      Yang, Xiaocui  and
      Wang, Daling  and
      Feng, Shi  and
      Chen, Yuxin  and
      Wang, Bixuan  and
      Zhang, Yifei",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.1192/",
    doi = "10.18653/v1/2025.findings-acl.1192",
    pages = "23221--23235",
    ISBN = "979-8-89176-256-5",
    abstract = "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](https://github.com/sci-m-wang/AnnaAgent)."
}