Mac Mini OpenClaw Memory: 8GB, Swap & Unified RAM (2026)

A base Mac mini with 8 GB or 16 GB unified memory is a popular home lab for 24/7 OpenClaw gateways — family WhatsApp bots, cron alerts, or a solo founder’s always-on assistant. Linux tutorials rarely mention macOS memory compression, swap on APFS, or Apple Silicon unified memory where the Gateway (Node.js) and Ollama / MLX local models fight over the same pool.

Most “OpenClaw won’t start” threads are not CPU limits — they are RAM budgeting problems: the Gateway baseline alone can reach ~400–700+ MB RSS on a cold start after recent releases (see OpenClaw gateway RSS discussion), before you load plugins, sessions, or a 7B local model.

This guide is for Apple ecosystem builders — not datacenter Xeon racks. You will learn how to measure, cap, and schedule memory on M1/M2/M3/M4 Mac mini hosts, with links to our OpenClaw multi-agent routing guide (why extra agents cost RAM), DeepSeek-R1 / local quantization, and remote SSH ops. Hardware trends: Computex 2026 agent infrastructure.

Introduction

This article covers unified memory budgeting, RAM tier choices, an eight-step macOS runbook, Activity Monitor signals, troubleshooting, and six FAQs for always-on OpenClaw.

Unified memory: one pool, two hungry workloads

Apple documents unified memory at Apple Silicon overview — CPU, GPU, and Neural Engine access the same bytes without discrete VRAM.

┌─────────────────────────────────────────┐
│     Apple Silicon unified memory         │
│  ┌─────────────┐    ┌─────────────────┐ │
│  │ OpenClaw    │    │ Ollama / MLX    │ │
│  │ Gateway     │◀──▶│ local LLM       │ │
│  │ (Node V8)   │    │ (model weights) │ │
│  └─────────────┘    └─────────────────┘ │
│           macOS compression + swap         │
└─────────────────────────────────────────┘

RAM tiers: what is realistic on 8 GB vs 16 GB

• If RAM is 8 GB → run Gateway 24/7, push inference to cloud API or one small Ollama model — not both heavy.
• If RAM is 16 GB → 7B Q4 + single Gateway is the sweet spot for Apple Silicon homelab.
• If you need 3+ isolated agents → upgrade RAM or split Gateway and Ollama across two Mac minis (SSH to second box).

Step-by-step runbook (macOS)

Step 1 — Baseline memory pressure

Open Activity Monitor → Memory. Note Memory Pressure (green/yellow/red) and Swap Used.

vm_stat | head -12
sysctl vm.swapusage
memory_pressure 2>/dev/null || true

Yellow/red pressure means macOS is already compressing or swapping — fix before adding OpenClaw.

Step 2 — Measure OpenClaw Gateway RSS

ps aux | grep -i '[o]penclaw' | awk '{print $2,$4,$6/1024"MB",$11}'

Community baselines: ~350 MB on older builds, 700 MB+ on some 2026.4.x cold starts. If RSS climbs daily, schedule a weekly restart until you trim plugins.

Step 3 — Cap Node.js heap (wrapper)

Direct NODE_OPTIONS in LaunchAgent plists may be stripped on regen. Use the wrapper from OpenClaw gateway install docs:

cat > ~/bin/openclaw-memwrap.sh << 'EOF'
#!/bin/bash
export NODE_OPTIONS="--max-old-space-size=2048"
exec /opt/homebrew/bin/openclaw "$@"
EOF
chmod +x ~/bin/openclaw-memwrap.sh

OPENCLAW_WRAPPER="$HOME/bin/openclaw-memwrap.sh" openclaw gateway install --force
openclaw doctor

On 8 GB use 1536–2048 MB heap cap; on 16 GB try 3072–4096 MB. Track OOMs in ~/Library/Logs/ or Console.app.

Step 4 — Run lean OpenClaw config

Edit ~/.openclaw/openclaw.json:

• One agents.list entry until RAM upgrade
• Disable unused channels and cron jobs
• Postpone memory search / dreaming plugins on 8 GB (memory config)
• Keep agent-to-agent off unless required (multi-agent docs)

openclaw gateway restart

Step 5 — Size Ollama / local models to unified memory

ollama pull llama3.2:3b-instruct-q4_K_M
ollama ps

Point OpenClaw model config to API or local endpoint per OpenClaw docs. Do not load 13B+ on 8 GB while Gateway stays resident — macOS will swap and Telegram webhooks will lag. See DeepSeek-R1 local quantization.

Step 6 — Swap and SSD hygiene (compression-friendly)

• Keep ≥ 20 GB free on the boot volume (swap lives on APFS)
• Prefer wired Ethernet for headless mini — Wi‑Fi debug menus steal RAM on GUI sessions
• Avoid Time Machine snapshots filling disk during agent peaks
• On mainland homelab links, stabilize egress so retries do not multiply Gateway buffers (unstable export bandwidth amplifies memory spikes)

while sleep 60; do date; sysctl vm.swapusage; done

Step 7 — Limit concurrent work (Node is not a thread pool)

launchctl setenv OLLAMA_NUM_PARALLEL 1

Step 8 — Headless 24/7 operations

• System Settings → Energy → prevent sleep on power adapter
• Use SSH for admin, not always-on VNC (SSH guide)
• openclaw gateway stop --disable before major macOS upgrades, re-enable after

Activity Monitor cheat sheet

Troubleshooting

JavaScript heap out of memory

Pattern: Gateway exits with FATAL ERROR: Reached heap limit. Fix: Lower concurrent cron; apply wrapper --max-old-space-size; upgrade to 16 GB; move Ollama to second machine.

Telegram/WhatsApp replies lag but CPU low

Pattern: vm.swapusage increasing, Memory Pressure yellow. Fix: Stop Ollama when using cloud models; use 3B quant; reduce Gateway plugins; reboot mini weekly.

Ollama loads then Gateway killed

Pattern: macOS memory compression maxed, processes jetsam. Fix: Serialize workloads — Gateway API-only during day, batch local inference at night; never run 7B + 3 agents on 8 GB.

Gateway RSS grows without traffic

Pattern: Known regression class on some versions — plugin scan at startup. Fix: Pin stable OpenClaw version; monitor RSS issue tracker; restart Gateway on cron.

FAQ

Conclusion

Mac Mini OpenClaw memory optimization is unified-memory budgeting: cap the Node Gateway, right-size local LLMs, keep swap-visible SSD headroom, and resist multi-agent sprawl on 8 GB. 16 GB is the practical floor for 7B Q4 + 24/7 Gateway; 8 GB remains viable for API-driven agents with a tight config.

Measure with vm_stat and Activity Monitor, install Gateway with a memory wrapper, and align software depth with our OpenClaw routing and quantization guides.

Official: OpenClaw Gateway CLI · GitHub openclaw/openclaw.