[BUG] Memory leak + Swap growth — Claude Code Sonnet 4.6 kills sessions in minutes

Date: 2026-02-21
Reporter: Linux user, collected via automated SRE monitoring
Claude Code version: 2.1.49
Severity: Critical — tool is barely usable; sessions die every 8–13 minutes

Environment

Observed Behavior

Claude Code's memory usage (RSS) grows continuously and rapidly from ~300 MB at launch to the MemoryMax ceiling within 8–13 minutes. The process is then killed by either:

1. Kernel OOM killer (memcg constraint) — when RSS hits MemoryMax
2. systemd-oomd — when userspace memory pressure exceeds 50–86% for >20s

Additionally, the kernel OOM reports ~7 GB of swap entries (swapents) in use at kill time — far beyond what a CLI tool should require. This indicates a concurrent RAM leak and swap growth issue.

Sessions are too short to be productive. The tool restarts ~8–10 times in 2 hours.

Crash Timeline — 2026-02-21 (today)

All kills confirmed from journalctl kernel + systemd-oomd logs:

Total kills on 2026-02-21: 9 sessions in ~2 hours
Average session duration: ~8.3 minutes

Raw Kernel OOM Events

12:37:53 kernel: Memory cgroup out of memory: Killed process 9867 (claude)
  total-vm:74926044kB, anon-rss:6095352kB, file-rss:37376kB, swapents:1819904

13:09:53 kernel: Memory cgroup out of memory: Killed process 12718 (claude)
  total-vm:74870336kB, anon-rss:6114484kB, file-rss:37760kB, swapents:1825472

13:19:50 kernel: Memory cgroup out of memory: Killed process 13271 (claude)
  total-vm:74878428kB, anon-rss:6169088kB, file-rss:49792kB, swapents:1826943

13:29:25 kernel: Memory cgroup out of memory: Killed process 13688 (claude)
  total-vm:74906228kB, anon-rss:6108972kB, file-rss:49792kB, swapents:1827432

13:42:31 kernel: Memory cgroup out of memory: Killed process 14519 (claude)
  total-vm:74845688kB, anon-rss:5967064kB, file-rss:49408kB, swapents:1831360

Notable:

• total_vm consistently ~71–75 GB (virtual address space — massive for a CLI tool)
• anon-rss consistently hits ~5.97–6.17 GB before kill
• swapents consistently ~1.82–1.83 million pages = ~7 GB of swap used per session

Live Monitoring Data

Monitoring started at 14:20:56 on new session (MemoryMax=10 GB).
Data collected every 30s into /tmp/claude-memory-log.csv:

timestamp,pid,VmRSS_MB,VmSwap_MB,MemCurrent_MB,RAM_free_MB,Swap_free_MB,MemMax_MB

(Log file at /tmp/claude-memory-log.csv — populated during 10-min background monitoring)

Snapshot at session start (14:20):

• MemoryCurrent = 1,537 MB (fresh session)
• MemoryMax = 10,240 MB (10 GB)
• RAM free = ~3,700 MB
• Swap used = ~922 MB (system-wide)

Growth Rate Analysis

From OOM kernel data and session durations:

systemd-oomd Events

12:55:30 systemd-oomd: Killed scope (memory pressure 56.06% > 50% for >20s)
13:55:09 systemd-oomd: Killed scope (memory pressure 86.47% > 50% for >20s)
13:55:24 systemd-oomd: Killed scope (memory pressure 82.38% > 50% for >20s)
13:55:40 systemd-oomd: Killed scope (memory pressure 73.54% > 50% for >20s)
14:07:14 systemd-oomd: Killed scope (memory pressure 73.36% > 50% for >20s)
14:16:11 systemd-oomd: Killed scope (memory pressure 73.60% > 50% for >20s)
14:16:26 systemd-oomd: Killed scope (memory pressure 73.57% > 50% for >20s)
14:16:41 systemd-oomd: Killed scope (memory pressure 66.35% > 50% for >20s)
14:16:56 systemd-oomd: Killed scope (memory pressure 63.59% > 50% for >20s)

Sessions were also killed via systemd-oomd before hitting MemoryMax, suggesting global system memory pressure caused by claude's leak.

Earlier Events (overnight 2026-02-21 00:00–02:00)

01:17:56 systemd-oomd: Killed terminal scope (pressure 89.18% > 50%)
01:18:11 systemd-oomd: Killed terminal scope (pressure 82.72% > 50%)
11:38:47 systemd-oomd: Killed terminal scope (pressure 65.87% > 50%)

Terminal processes were collateral damage from claude's memory pressure.

Current Workaround (Insufficient)

# In ~/.bashrc:
alias claude='systemd-run --user --scope -p MemoryMax=10G claude'

• At 6 GB limit: sessions lasted ~8–13 min before kill
• At 10 GB limit: expected improvement is marginal given ~630 MB/min growth rate
  • Estimated session duration at 10 GB: ~15–18 min (still too short)
  • At 14 GB: ~20–25 min
  • There is no MemoryMax value that fixes a leak growing at 630 MB/min

Impact

• Workflow completely disrupted — cannot complete any task longer than ~8 minutes
• Sessions killed mid-operation, leaving incomplete work
• System-wide memory pressure kills other processes (terminal, tracker-miner) as collateral
• User must constantly restart sessions and re-explain context
• 9+ forced restarts in a single 2-hour work session

Suspected Root Cause

Based on the data:

1. Heap/context buffer leak: total_vm of ~71 GB and continuous anonymous RSS growth suggests unbounded heap or context window buffer accumulation
2. Swap pre-allocation: swapents reaching ~7 GB (nearly the full 8 GB swapfile) before kill suggests swap is being aggressively mapped/not released between operations
3. No GC or memory cap: The process does not appear to release memory between API calls or conversation turns

This regression appears to have started around the Sonnet 4.6 release (~Feb 15, 2026). Earlier versions did not exhibit this behavior at this severity.

Explicit Request

1. Fix the RAM memory leak in Claude Code 2.1.49 / Sonnet 4.6 — unbounded RSS growth must be addressed
2. Fix the swap growth issue — swapents reaching ~7 GB per session is abnormal
3. Add internal memory management — GC, context trimming, or soft memory cap before process termination
4. Provide a --memory-limit flag as interim workaround allowing graceful degradation instead of hard kill

Reproduction

1. Install Claude Code 2.1.49 on Linux with 16 GB RAM + 8 GB swap
2. Run: systemd-run --user --scope -p MemoryMax=6G claude
3. Start any multi-turn conversation or coding task
4. Monitor: watch -n5 "systemctl --user show \$(systemctl --user list-units | grep run- | grep scope | awk '{print \$1}' | head -1) | grep -E 'MemoryCurrent|MemoryMax'"
5. Observe: MemoryCurrent grows ~630 MB/min until process is killed

Bug report generated automatically via SRE monitoring script on 2026-02-21.