chatdev-memorysync

MemorySync memory store for ChatDev 2.0 (DevAll) — OpenBMB's zero-code multi-agent platform.

One pip install adds type: memorysync to ChatDev's memory registry: usable from workflow YAML and the web console's store-type dropdown, no code changes to ChatDev required.

# inside your ChatDev checkout's environment
uv pip install chatdev-memorysync     # or: pip install chatdev-memorysync

Activation is automatic (a .pth hook registers the store whenever ChatDev's modules are importable, and silently no-ops anywhere else). The explicit alternative:

import chatdev_memorysync
chatdev_memorysync.register()

Workflow YAML

graph:
  nodes:
    - id: writer
      type: agent
      config:
        # ... model config ...
        memories:
          - name: user_memory
            top_k: 5
            retrieve_stage: [gen]
            read: true
            write: true
  memory:
    - name: user_memory
      type: memorysync
      config:
        api_key: ${MEMORYSYNC_API_KEY}
        user_id: customer-42        # REQUIRED — who these memories belong to
        session_id: support         # optional session label

Retrieved memories arrive in the standard ===== Related Memories ===== block ChatDev injects into agent prompts.

What gets stored

After each agent node that writes to the store (write: true, ChatDev's default) runs, the store sends that node's input — the request it received, with ChatDev's === INPUT FROM … === pipeline headers stripped — to MemorySync as the user's turn. MemorySync extracts the durable facts in it (preferences, constraints, details about the user) and only those facts become memories. The input text itself is not stored.

• Agent outputs are not stored as memories, so they are never sent. store_outputs is still accepted in workflow YAML and has no effect.
• A retry or re-run of the same input in the same session is recognised server-side and extracted once.
• Facts carry surface: chatdev, your session_id, and role: input. The input's attachment overview travels with the request, but facts keep only plain (scalar) metadata, so attachment lists are not stored on them.
• Inputs under 3 characters are skipped; inputs over 16,000 characters are truncated before they are sent.

A downstream node's input is what the upstream nodes produced, and it is sent the same way. To learn only from the user's own words, keep write: true on the node that receives the user's request and set write: false on the others (they can still read).

Why this instead of the in-tree Mem0 store?

ChatDev ships a mem0 store in its repo. Verified against its source:

Facts extracted here are recallable from every other MemorySync surface (LangChain, the CLI, voice agents, Dify…) and vice versa.

Configuration

Per-agent knobs (top_k, similarity_threshold, read, write, retrieve_stage) come from ChatDev's own memories: attachment config and are honored — including similarity_threshold filtering on real scores.

Tests

git clone --depth 1 https://github.com/OpenBMB/ChatDev.git chatdev
pip install -e . pytest pydantic openai pyyaml tenacity numpy faiss-cpu fastmcp jinja2
pytest tests -q   # 38 checks against the real ChatDev 2.0 modules

The suite drives ChatDev's real config parser, registry, MemoryFactory, and MemoryManager — and reproduces each in-tree Mem0 store bug as a regression test.

License

MIT © MemorySync.