Mission#

Write down everything future-me needs to know to operate, troubleshoot, and extend the local LLM setup. Goldfish memory rule: if it's not written down, it doesn't exist.

Deliverable#

A new ## Local LLM section in MEMORY.md with:

• Where Ollama runs (host, port, service name)
• How to start/stop/restart it (systemctl --user ...)
• How to add a new model (ollama pull, then update agents.defaults.models in config)
• The provider config block (full JSON) so I can re-create it from scratch
• VRAM footprint of each currently-installed model
• Which agent tasks/cron jobs currently route to local models
• Common failure modes + how to diagnose (Ollama down, model not loaded, OOM)

Plus a runbook at ~/.openclaw/workspace/projects/local-llm/RUNBOOK.md:

• "Ollama is down" — checklist
• "Local responses are slow / weird" — checklist
• "I want to add a new model" — recipe
• "I want to route a new task to local" — recipe
• "I'm switching to a different machine" — what to copy

Acceptance criteria#

• MEMORY.md has a complete Local LLM section
• RUNBOOK.md exists with the four scenarios above
• At least one other agent (Debbie or Chris) can read the docs and operate Ollama without asking Maxime