Mem0 memory in NAT workflows

datarobot-genai includes a NAT MemoryEditor provider that adapts the DataRobot Mem0 client for NAT's auto_memory_agent. This avoids the upstream nvidia-nat-mem0ai plugin while still using NAT's standard memory: configuration.

Install the NAT extra (includes mem0ai):

pip install "datarobot-genai[nat]"

Set the Mem0 API key at runtime:

export MEM0_API_KEY=...

Configure the memory provider

memory:
  mem0_memory:
    _type: dr_mem0_memory
    # Optional explicit override; defaults from MEM0_API_KEY runtime settings.
    # api_key: ${MEM0_API_KEY}
    # Optional Mem0 organization/project routing:
    # host: https://api.mem0.ai
    # org_id: ...
    # project_id: ...

mem0_memory is the local name you reference from the workflow. dr_mem0_memory is the DataRobot provider type registered by this package. The provider reads MEM0_API_KEY through DataRobot app framework settings by default; use api_key only when you need an explicit workflow-level override.

Wrap an agent with automatic memory

middleware:
  datarobot_dragent_normalization:
    _type: datarobot_dragent_normalization

functions:
  nat_agent:
    _type: per_user_tool_calling_agent
    llm_name: datarobot_llm
    # Required for inner agents: NAT middleware is per-function and opt-in — it is
    # NOT inherited from the parent workflow. The memory wrapper calls the inner
    # agent's stream directly, so the normalization middleware must be declared here
    # for the inner agent to yield DRAgentEventResponse.
    middleware:
      - datarobot_dragent_normalization
    tool_names:
      - planner
      - writer

workflow:
  _type: auto_memory_agent
  inner_agent_name: nat_agent
  memory_name: mem0_memory
  llm_name: datarobot_llm
  description: "Agent with automatic memory capture and retrieval."

per_user_tool_calling_agent emits native NAT output (a str for single responses and ChatResponseChunk for streaming). The datarobot_dragent_normalization middleware converts that into DRAgent's canonical DRAgentEventResponse. When a per_user_tool_calling_agent is the top-level workflow (not wrapped by a memory agent), declare the middleware in the workflow: block instead — as the last (innermost) entry so it runs before any moderation or telemetry middleware.

NAT supplies the runtime user_id to the memory backend from the session user manager, X-User-ID header, or its local fallback. The provider forwards that user_id into Mem0 v2 search filters so memories remain isolated per user.

Optional search and add parameters

Parameters under search_params are passed to MemoryEditor.search(), and parameters under add_params are passed to MemoryEditor.add_items():

workflow:
  _type: auto_memory_agent
  inner_agent_name: nat_agent
  memory_name: mem0_memory
  llm_name: datarobot_llm
  search_params:
    top_k: 5
  add_params:
    agent_id: blog_agent

The provider also supports host, org_id, and project_id on the memory config for Mem0 deployment routing.