NVIDIA NemoClaw Memory-Driven Agent Splits Evidence From Judgment

NVIDIA published a memory-driven Chief of Staff agent built with NemoClaw that keeps a structured self model of people, projects, priorities, and working patterns. The design separates evidence from judgment: Markdown pages store derived knowledge, while a SQLite ledger records obligations, rankings, corrections, and audit events. This split helps developers trace whether a wrong answer came from source material, memory maintenance, retrieval, or the model's final decision.

Intent gate and correction loop

An intent gate reserves the highest priority tier for obligations tied to stated user priorities, not short-term urgency. Deterministic code enforces tier size, overflow behavior, and ranking order. Users can move or ignore obligations, and each change lands in an append-only audit trail. Repeated correction patterns update a readable preference policy that stays inspectable and editable.

Benchmark results

On the Agent Memory Benchmark, the self model improved overall accuracy to 90.9% from 82.8% for an agentic RAG baseline. Hard questions rose to 87.1% from 67.7%, tracking of changed facts went to 100% from 60.0%, and point-in-time reasoning doubled to 66.7% from 33.3%. Entity disambiguation improved to 86.7% from 66.7%, and multisource synthesis to 94.5% from 87.7%. Two categories dipped: answered faithfully based on corpus fell to 92.3% from 100%, and single-hop lookup to 83.3% from 86.7%. Citation coverage rose to 97.8% from 92.5%. Both configurations used NVIDIA Nemotron 3 Ultra.

Runtime boundaries

NemoClaw integrates the example with NVIDIA OpenShell, which sandboxes the agent and governs file system, process, and network access. Credentials for managed inference and MCP connections stay outside the sandbox. Memory and retrieved content are inputs, not trusted policy, so a misinterpreted instruction still operates within operator-defined limits.

The recipe packages the example as a deployable Hermes profile with a structured memory schema, a durable obligation ledger, bounded ranking logic, user correction and audit paths, scheduled memory maintenance, synthetic messages and memory pages, an offline walkthrough, and unit tests. The current recipe does not send messages or modify source systems, so it can be examined without connecting a workplace account. Live connectors require separate handling for credentials, privacy, retention, and deletion.