openclaw - 💡(How to fix) Fix Critical: Gateway Memory Leak — RSS grows from 350MB to 15.5GB over days, causing repeated OOM crashes

Summary

The OpenClaw gateway process suffers from a severe memory leak. RSS grows from ~350 MB at startup to 15.5 GB over 2-3 days of normal use, eventually causing the process to be killed by the OS OOM killer. This triggers repeated launchd-handoff restart cycles.

Environment

• OpenClaw version: 2026.6.1 (2e08f0f)
• Node.js: v22.22.0
• OS: macOS 20.6.0 (Big Sur), x64
• System RAM: 8 GB
• Plugins loaded: claw-mem (v6.15.0), claw-ctx (v5.2.0) — custom local plugins

Evidence

Before Restart (memory leak peak)

PID 19893  RSS: 15,513 MB  CPU: 89.7%  (openclaw-gateway)
PID 40226  RSS:    201 MB  CPU:  0.0%  (openclaw-tui, running 2+ days)

After Restart (immediately)

PID 19926  RSS:    350 MB  CPU:  0.3%  (openclaw-gateway)
PID 40226  RSS:    197 MB  CPU:  0.0%  (openclaw-tui)

The gateway dropped from 15.5 GB to 350 MB after a single restart, confirming a memory leak in the long-running process, not a configuration or disk-space issue.

Restart/Crash Frequency

From ~/.openclaw/logs/gateway-restart.log, launchd-handoff restarts (process died → launchd auto-restarted):

Crashes Produce No Error Logs

The gateway error log (gateway.err.log) has no entries since May 21. The process dies silently — consistent with OOM kill or Node.js heap exhaustion with no chance to write diagnostics.

Analysis

1. Memory growth is monotonic: RSS increases steadily over days, never decreases.
2. No automatic recovery: The process dies silently, leaving only launchd to restart it.
3. Plugin state accumulation suspected: With claw-mem (in-memory index, search cache) and claw-ctx (token caches, session state, drift alerts) loaded, the gateway never purges accumulated state between sessions.
4. No heap limit configured: The gateway Node.js process has no --max-old-space-size flag. On an 8 GB machine, this means it can consume all available RAM before being killed.
5. Physical RAM saturation: 15.5 GB RSS on an 8 GB machine causes heavy swapping, degrading the entire system before the OOM kill.

Impact

• Critical: Gateway becomes unresponsive or crashes multiple times per day
• Intermediate crashes interrupt ongoing conversations and tool operations
• Repeated restart cycles generate noise in stability logs
• Degrades confidence in OpenClaw as a reliable long-running agent platform

Suggested Fixes

1. Root cause: Investigate and fix memory leaks in the gateway process, particularly around plugin lifecycle (plugin state not released between sessions) and internal caches (unbounded growth).
2. Self-protection: Add a --max-old-space-size default (e.g., 4096 MB) to the gateway Node.js process to trigger a clean GC/OOM error before the OS OOM killer.
3. Health monitoring: Implement an RSS-based self-restart (e.g., if RSS > 4 GB, gracefully restart). This is a pragmatic safeguard even after the leak is fixed.
4. Periodic restart: Consider an optional --auto-restart-interval flag (e.g., daily at 3 AM) for production deployments.

Steps to Reproduce

1. Start openclaw gateway with memory/caching plugins loaded
2. Use the gateway normally over 2-3 days (send messages, perform tool calls)
3. Monitor RSS with ps -o pid,rss,comm
4. Observe RSS grow from ~350 MB to 10+ GB
5. Eventually the process is killed by the OS

Additional Context

A prior separate issue (May 21) caused a startup crash loop: the claw-mem plugin manifest was not found at the configured path. This was fixed by rebuilding the plugin. The two issues are independent but compound: the startup crash prevents recovery after the OOM kill.

Submitted via gh CLI by Jarvis (Peter Cheng's assistant)