NVIDIA NemoClaw-Based Memory Agent Separates Knowledge and Judgment, Improving Accuracy from 82.8% to 90.9%

On September 4, 2026, NVIDIA released a 'Memory-Driven Chief of Staff' agent based on NemoClaw. This system structurally separates knowledge (facts) and judgment (opinions) to overcome the limitations of conversation logs or RAG, raising benchmark accuracy from the previous 82.8% to 90.9%.

Architecture and Core Technology

The agent stores knowledge in a Markdown wiki, records judgments in SQLite, and performs execution control via the Rust-based OpenShell. In particular, under the principle that 'context cannot authorize actions,' memory is designed to only make recommendations, while actual actions require credentials and user approval. The self-model stores original text and derived interpretations separately to ensure ease of debugging.

Performance and Cost Analysis

In benchmarks based on Nemotron 3 Ultra, temporal metrics (chain_freshness +40%p, as_of +33.3%p) improved significantly. However, challenges remain in cost efficiency, such as the consumption of 183 million tokens during the collection stage. Token usage per question was also higher than the baseline, indicating that the argument for 'savings after investing in collection costs' does not hold.

Deployment and Licensing

The current release supports only Slack and Outlook connectors and is at the level of a reference recipe rather than a production environment. NemoClaw, OpenShell, and the benchmark dataset are released under the Apache License 2.0, allowing for research and commercial use.