NVIDIA AI Updates: September 5, 2026

1. NemoClaw Chief of Staff Agent Separates Evidence, Knowledge, and Execution

NVIDIA. NVIDIA published a reference build for a memory-driven agent on NemoClaw that maintains a “self model”, a human-readable Markdown knowledge layer covering people, projects, priorities, goals, and recurring work patterns with defined schemas for indexing and cross-references. The architecture splits into three layers: raw evidence from daily work, structured knowledge in Markdown, and governed execution, with a SQLite ledger recording obligations, rankings, corrections, and audit events so evidence stays separate from judgments. Intent gating ranks stated priorities above short-term urgency, and append-only audit trails turn user corrections into readable preference policies. NVIDIA OpenShell supplies runtime sandboxing over file system, process, and network access. On the Agent Memory Benchmark against Nemotron 3 Ultra, overall accuracy rose from 82.8% to 90.9%, tracking changed facts went from 60.0% to 100.0%, and entity disambiguation went from 66.7% to 86.7%. Source

2. NVFP4 Plus Speculative Decoding Puts Reasoning Models on Jetson

NVIDIA. A developer blog walks through deploying reasoning models locally on Jetson AGX Thor, Jetson AGX Orin, and Jetson Orin Nano, arguing the size constraint that forced agent builders to route inference through a data center is lifting. Combining NVFP4 quantization with speculative decoding delivers 3.37x over a BF16 baseline for Nemotron 3.5 Lightning, a 30B mixture-of-experts model activating 3B parameters per token, and 6.28x for the dense Qwen3.8-27B. Measured throughput lands at 123 to 138 output tokens per second for Nemotron and 27.69 to 34.44 for Qwen3.8-27B. The best draft method is model-specific: DSpark for Nemotron 3.5 Lightning, DFlash2 for Qwen3.8-27B. Gemma 4 E4B is the suggested starting point on Orin Nano. NVIDIA cautions that general benchmarks cannot confirm a quantized checkpoint preserves behavior critical to a specific dataset, so application-level validation remains necessary. Source