Problem

Installing openclaw globally pulls in node-llama-cpp (3.15.1) as a hard dependency, which installs ~670MB of pre-compiled binaries for every supported platform and GPU backend:

This is especially wasteful in environments like GitHub Codespaces or CI/CD where:

• No GPU is available (CUDA/Vulkan binaries are useless)
• Only one platform is needed (ARM binaries on x64 are unused)
• Local LLM embedding (memory search) is not used — cloud APIs handle inference

For our team, this means 670MB of dead weight per codespace that we currently work around by deleting the binaries after install in our Dockerfile.

Proposal

Make node-llama-cpp an optional dependency (or move it to a plugin/extension) so users who don't need local LLM embedding can skip it entirely.

Suggested approaches (pick any):

1. Move node-llama-cpp from dependencies to optionalDependencies in package.json
2. Extract local embedding into a separate package (e.g., @openclaw/memory-local)
3. Use a lazy-install pattern — download llama binaries on first use, not at install time

Current workaround

# After npm install -g openclaw@latest:
RUN rm -rf /usr/local/share/npm-global/lib/node_modules/openclaw/node_modules/@node-llama-cpp/linux-x64-cuda-ext \
           /usr/local/share/npm-global/lib/node_modules/openclaw/node_modules/@node-llama-cpp/linux-x64-cuda \
           /usr/local/share/npm-global/lib/node_modules/openclaw/node_modules/@node-llama-cpp/linux-x64-vulkan \
           /usr/local/share/npm-global/lib/node_modules/openclaw/node_modules/@node-llama-cpp/linux-arm*

Environment

• openclaw@2026.2.15
• node-llama-cpp@3.15.1
• GitHub Codespaces (linux-x64, no GPU)
• 32GB disk — 670MB is ~2% of total disk