OOM kill: Claude Code memory grows to 10-13 GB after just 3 exchanges

Summary

Claude Code's Node.js process grows to 10-13 GB RSS after only ~3 conversation exchanges, triggering the Linux OOM killer. This is NOT from heavy/prolonged use — the session had barely started.

Environment

• Ubuntu 24.04, kernel 6.17.0-14-generic
• 16 GB RAM, 4 GB swap
• Claude Code CLI (latest)
• Alienware x14
• Connected to hotel WiFi in China, session topic was VPN/WiFi connectivity debugging

What happened

1. Started a Claude Code session to debug WiFi and VPN connectivity (hotel WiFi in China)
2. After only ~3 exchanges, the process memory grew rapidly to 10+ GB RSS
3. Linux OOM killer terminated the process
4. Restarted Claude Code, same thing happened — killed again after a few exchanges
5. This happened 3 times in a row

Context

• This has never happened before — the user regularly runs 5-6 concurrent Claude Code sessions doing complex queries and database work without memory issues
• The only difference was the topic (China VPN/WiFi connectivity) and being on hotel WiFi
• Normal conversational use stays at ~200-400 MB
• The session involved basic system commands (checking network config, systemctl status, etc.) — nothing unusual

Kernel OOM logs

oom-kill: task=claude, pid=4694, uid=1000
Out of memory: Killed process 4694 (claude) total-vm:74797396kB, anon-rss:11151744kB (~10.6 GB)

oom-kill: task=claude, pid=5712, uid=1000
Out of memory: Killed process 5712 (claude) total-vm:76326728kB, anon-rss:13408196kB (~12.8 GB)

Both kills happened within the same hour. oom_score_adj:200 on the claude process.

Expected behavior

Claude Code should not consume 10+ GB RAM after 3 exchanges regardless of topic. There should be a reasonable memory ceiling — either via Node.js --max-old-space-size or internal memory management.

Possible factors

• Network conditions (hotel WiFi in China, possibly slow/unstable connection)
• Could be related to retries or buffering when network is degraded
• VPN-related system commands in the conversation (but output was minimal)