Automatic Memory Wrapper for NeMo Agent Toolkit Agents

Why it matters

Memory that depends on the model remembering to call a tool is captured inconsistently. Memory that the runtime captures has to be scoped to a user, or it leaks between users.

What to practise

Configure an auto_memory_agent workflow that wraps a ReAct agent through inner_agent_name and memory_name, with Zep or Mem0 as the backend. Toggle the flags that save user messages, retrieve memory before every response, and save agent replies, and tune search_params and add_params. Run it with nat run --user_id and --conversation_id, and watch that memory operations fail closed when no identity is resolved. Read why a trusted identity_header is safe only behind an authenticating proxy.

How it fits

The wrapper routes conversation history into a long-term store; compare that choice with the store roles in 1.4.1. The user and conversation scoping, and the identity-header trust boundary, are 1.4.2’s authenticated memory scopes.