NemoClaw: agentic tool issues
Conversations since GTC describe NemoClaw as an ‘agentic’ AI tool being used to build autonomous assistants, but users report aggressive state-management that can cause memory bloat in production deployments. The issue was flagged in social posts noting the agentic push at GTC 2026 alongside memory-growth problems during long-running runs (x.com).
NVIDIA is pitching NemoClaw as a safer way to run always-on AI assistants, while users and developers are warning that its persistent state can swell over long runs. (investor.nvidia.com) NVIDIA announced NemoClaw on March 16, 2026 at GTC as a one-command stack for OpenClaw that installs Nemotron models and the OpenShell runtime for “self-evolving” agents. The company’s docs say NemoClaw handles onboarding, lifecycle management and OpenClaw operations inside OpenShell containers. (investor.nvidia.com) (docs.nvidia.com) In plain terms, NemoClaw wraps an autonomous assistant in a sandbox and gives it files that persist across sessions, the way a notebook carries forward instructions and memories. NVIDIA’s workspace docs list files such as SOUL.md, USER.md, IDENTITY.md, AGENTS.md, MEMORY.md and daily memory notes, and say the agent reads them at the start of every session. (docs.nvidia.com) (docs.openclaw.ai) That design is where the production complaint starts. If an agent keeps appending rules, summaries and notes to files that are reloaded every session, the “notebook” can grow until memory use, token use or both become harder to control in long-running deployments. (docs.nvidia.com) (github.com) NemoClaw’s own materials emphasize “state management” as a feature, not a side effect. The GitHub repository says NemoClaw adds state management on top of OpenShell, and NVIDIA’s backup docs say those workspace files persist across sandbox restarts unless the sandbox is explicitly destroyed. (github.com) (docs.nvidia.com) The same persistence that helps an assistant remember a user can also preserve bad data, stale instructions or secrets. One GitHub issue opened in early April warned that API keys or tokens written into persistent memory files could survive across sessions and backups, turning workspace memory into a long-lived exfiltration surface. (github.com) Other community threads show the tradeoff from a different angle. A discussion updated on April 9 included “real-world memory measurements” for NemoClaw on DGX Spark, while another issue from early April asked for backup and restore hooks after workspace data was lost on pod restart despite persistent storage. (github.com 1) (github.com 2) NVIDIA is also signaling that the project is still early. The build page says NemoClaw and OpenShell are in “early preview,” and the troubleshooting guide tells users to file issues or ask for help in Discord if their problem is not documented. (build.nvidia.com) (docs.nvidia.com) The current split is straightforward: NVIDIA is selling guardrails, local models and easier deployment for autonomous assistants, while operators are discovering that durable memory needs cleanup rules as much as it needs persistence. NemoClaw’s next test is not the GTC demo, but whether long-running agents can keep state without letting that state keep growing. (nvidia.com) (github.com)
