Agent Skills: Memory Guardian Skill

Memory Guardian Skill

Purpose

Monitors WSL2 memory usage and enforces safe scaling limits for Claude Code instances and parallel agents.

When This Skill Activates

• Before spawning 3+ parallel Task agents
• Before running npm run build or other memory-heavy operations
• When SessionStart hook reports WARNING or CRITICAL memory status
• When user mentions running multiple CC instances

Memory Check Command

free -m | awk '/^Mem:/{printf "RAM: %dMB/%dMB (%d%%)\n", $3, $2, $3*100/$2} /^Swap:/{if($3>0) printf "Swap: %dMB/%dMB\n", $3, $2}'

Scaling Rules

Green Zone (< 75% RAM usage)

• Parallel Task agents: up to 4
• Background builds: safe
• No restrictions

Yellow Zone (75-89% RAM usage)

• Parallel Task agents: max 2
• Prefer sequential over parallel work
• Avoid npm run build alongside agents
• Warn user: "Memory pressure is high. Limiting parallel agents to 2."

Red Zone (>= 90% RAM usage)

• Parallel Task agents: max 1 (or 0 — work sequentially)
• NO background builds
• Suggest user close other CC instances
• Warn user: "Memory critically low. Working sequentially to prevent WSL crash."

Before Heavy Operations

Always run the memory check before:

1. Spawning 3+ parallel Task subagents
2. Running production builds (npm run build)
3. Starting dev server + agents simultaneously

WSL Configuration Reference

• Host: 16GB RAM
• WSL limit: 12GB (C:\Users\Frank\.wslconfig)
• Swap: 12GB
• Per CC instance: ~500MB + ~100-150MB per MCP server
• Safe concurrent CC instances: 4-5
• autoMemoryReclaim: dropcache (aggressive)

Recovery If Memory Is Critical

# Check what's using memory
ps aux --sort=-%mem | head -20

# Kill any orphaned node processes from old CC sessions
pkill -f "node.*claude" 2>/dev/null

# Force kernel to drop caches (non-destructive)
echo 1 | sudo tee /proc/sys/vm/drop_caches 2>/dev/null

Integration

• Hook: .claude/hooks/memory-check.sh runs on SessionStart
• Statusline: Memory percentage visible if hook fires warning
• Self-enforcing: CC reads this skill and self-limits before scaling