ChatDev - Mem0

Build multi-agent workflows in ChatDev

ChatDev is a zero-code multi-agent platform where agents, tools, and workflows are defined entirely in YAML. Mem0 integrates as a built-in memory store (type: mem0), giving your agents cloud-managed semantic search and cross-session persistence, all without writing any code.

Overview

In this guide, you’ll:

1. Set up ChatDev with the Mem0 memory store
2. Configure agents with persistent memory using YAML
3. Enable automatic memory retrieval and storage across conversations
4. Leverage cross-session persistence for personalized multi-agent interactions

Prerequisites

• Python 3.12+
• uv: Python package manager
• Node.js 18+ and npm: only needed if using the web console
• A Mem0 API key from app.mem0.ai
• An OpenAI API key (or another LLM provider supported by ChatDev)

Setup and Configuration

Install ChatDev and its dependencies (includes mem0ai):

git clone https://github.com/OpenBMB/ChatDev.git
cd ChatDev
uv sync

If you plan to use the web console, also install the frontend:

cd frontend && npm install && cd ..

Set up your environment variables in a .env file:

Get your Mem0 API key from Mem0 Platform.

MEM0_API_KEY=your-mem0-api-key
API_KEY=your-openai-api-key
BASE_URL=https://api.openai.com/v1

Configure Mem0 Memory Store

In your ChatDev workflow YAML, add a Mem0 memory store in the memory section:

memory:
  - name: mem0_store
    type: mem0
    config:
      api_key: ${MEM0_API_KEY}
      user_id: my-user-123       # optional: scope memories to a user
      agent_id: my-agent         # optional: scope memories to an agent

Mem0 handles all storage, embeddings, and search server-side. No local vector databases or embedding models are needed.

Attach Memory to an Agent

Reference the memory store in your agent node’s memories list:

nodes:
  - id: writer
    type: agent
    config:
      role: |
        You are a knowledgeable writer. Use your memories to build
        on past interactions.
      memories:
        - name: mem0_store
          top_k: 5
          similarity_threshold: 0.5   # minimum relevance score (0.0–1.0); set to -1.0 to disable
          retrieve_stage:
            - gen
          read: true
          write: true

• read: true: Agent retrieves relevant memories before generating a response
• write: true: Agent stores new memories from user input after each interaction
• top_k: Number of memories to retrieve per query
• similarity_threshold: Minimum relevance score for retrieved memories. Set to -1.0 to return all results regardless of score
• retrieve_stage: When to retrieve memories. Options: pre_gen_thinking, gen, post_gen_thinking, finished

Full Example Workflow

Here’s a complete workflow YAML that creates a memory-backed conversational agent:

version: 0.4.0
graph:
  description: Memory-backed conversation using Mem0

nodes:
    - id: writer
      type: agent
      config:
        base_url: ${BASE_URL}
        api_key: ${API_KEY}
        provider: openai
        name: gpt-5.4
        role: |
          You are a knowledgeable writer. Use your memories to build
          on past interactions. If memory sections are provided
          (wrapped by ===== Related Memories =====), incorporate
          relevant context from those memories into your response.
        params:
          temperature: 0.7
          max_tokens: 2000
        memories:
          - name: mem0_store
            top_k: 5
            retrieve_stage:
              - gen
            read: true
            write: true

memory:
    - name: mem0_store
      type: mem0
      config:
        api_key: ${MEM0_API_KEY}
        user_id: project-user-123
        agent_id: writer-agent

start:
    - writer
  end: []

Run the workflow:

# Option 1: CLI (recommended for quick testing)
uv run python run.py --path yaml_instance/demo_mem0_memory.yaml --name my_project

# Option 2: Web console
make dev
# Backend starts at http://localhost:6400, frontend at http://localhost:5173

To use the web console, open http://localhost:5173, create a new workflow, and paste your YAML configuration into the editor. The web console provides a visual chat interface for interacting with your memory-backed agents.

How It Works

When an agent with Mem0 memory receives input, the following cycle runs automatically:
1. Retrieve: Before generating a response, ChatDev queries Mem0 with the user’s input using semantic search. Relevant memories are injected into the agent’s context in this format:

===== Related Memories =====
--- mem0_store ---
1. User's favorite language is Rust
2. User lives in San Francisco
===== End of Memory =====

2. Generate: The agent produces a response using the retrieved memories as additional context.
3. Store: After generation, the user’s input is sent to Mem0 via client.add(). Mem0’s extraction model automatically identifies and stores facts, preferences, and key information. Only user input is stored. Agent output is excluded to keep memories clean. Memories persist in Mem0’s cloud across all sessions. The next time the same user_id or agent_id is used, previous memories are automatically retrieved.

Dual-Scope Memory (User + Agent)

When both user_id and agent_id are configured, Mem0 uses an OR filter to search across both scopes in a single query:

memory:
  - name: shared_store
    type: mem0
    config:
      api_key: ${MEM0_API_KEY}
      user_id: alice              # stores user preferences
      agent_id: support-bot       # stores agent-learned context

This means retrieval returns memories from both the user’s scope and the agent’s scope. Writes include both IDs, so each memory is accessible from either dimension. Use this when you want an agent to remember both what the user told it and what the agent learned across sessions.

Configuration Reference

Memory Store Config

Memory Attachment Config

Tips and Common Pitfalls

Indexing delay: Freshly stored memories may take a few seconds to become searchable. If a memory isn’t retrieved immediately after being stored, wait a moment and try again.

• No memories returned on first run: This is expected. Memories are stored after the agent responds, so the first interaction has no prior context. Memories appear starting from the second interaction onward.
• mem0ai not installed: If you see ImportError: mem0ai is required for Mem0Memory, run uv add mem0ai or pip install mem0ai to add the dependency.
• Invalid API key: A wrong or expired MEM0_API_KEY will log errors like Mem0 search failed or Mem0 add failed but won’t crash the agent. Check your key at app.mem0.ai.
• Pipeline headers in memories: ChatDev automatically strips internal pipeline headers (e.g., === INPUT FROM TASK (user) ===) before sending text to Mem0, so your memories stay clean.
• Clearing test memories: To delete memories created during testing, use the Mem0 dashboard at app.mem0.ai or the Python SDK: MemoryClient().delete_all(user_id="your-test-user").

Key Features

1. Zero-Code Integration: Configure Mem0 entirely through YAML, no Python code required
2. Cloud-Managed Storage: Mem0 handles embeddings, persistence, and search server-side
3. Semantic Search: Retrieve contextually relevant memories, not just keyword matches
4. Cross-Session Persistence: Memories survive across runs, sessions, and restarts
5. Multi-Agent Memory Sharing: Multiple agents can share memories through common user_id or agent_id scopes
6. Intelligent Input Processing: Only user input is stored; agent output is excluded to prevent noisy memories

Conclusion

By adding Mem0 as a memory store in ChatDev, your multi-agent workflows gain persistent, intelligent memory with zero code changes. Agents automatically remember past interactions and use that context to provide personalized, coherent responses across sessions.