NemoClaw
NVIDIA NeMo Agent sandboxes that learn from every conversation.
--harness nemoclawWhat you get
NemoClaw runs an OpenClaw gateway inside an NVIDIA NeMo Agent sandbox. Without self-driving agents, every conversation in the sandbox starts fresh. With them installed, the agent learns from each chat — preferences, decisions, recurring patterns — and walks into the next session already informed.
- Learns from every conversation — preferences, decisions, and recurring patterns are captured automatically. The agent walks into the next session already knowing them.
- Builds its own playbooks — the agent maintains long-form pages (a personal wiki) that self-rewrite from what it has learned.
- Sandboxed by NeMo's policy engine — the CLI patches just the network rules the plugin needs to reach Hindsight; everything else stays locked under Landlock.
Setup
- 1
Have NemoClaw on PATH
The CLI shells out to
nemoclaw list,nemoclaw <sandbox> status, andnemoclaw <sandbox> skill install. Create a sandbox first vianemoclaw onboardif you don't have one. - 2
Install the agent
One CLI call sets up the openclaw plugin on the host, patches the sandbox network policy, configures the connection (interactive on first run), ingests seed content, and installs the agent-knowledge skill into the sandbox.
npx @vectorize-io/self-driving-agents install marketing/seo --harness nemoclaw - 3
Connect to the sandbox
Brings up the sandbox's gateway.
nemoclaw <sandbox> connect - 4
Open a session
Memory works automatically from here on.
openclaw tui --session agent:main:main:session1
How it works
NemoClaw runs OpenClaw inside an NVIDIA NeMo Agent sandbox with read-only config and Landlock filesystem isolation. The integration is the same @vectorize-io/hindsight-openclaw plugin (≥ 0.7.2) the OpenClaw harness uses, but configured through hindsight-nemoclaw setup so the sandbox's network policy gets patched in the right place. Two hooks fire on every turn: a pre-turn recall step injects an invisible <hindsight_memories> system block, and a post-turn retain step saves the conversation. The plugin also exposes the tools below.
Tools the agent gets
list_pagesList page ids and names.get_pageRead the content of a page.create_pageDeclare a new page recipe.update_pageChange a page's name or source query.delete_pageRemove a page.recallSearch retained memories.ingestAdd new text content as a memory.