hermes - 💡(How to fix) Fix hermes update restarts gateway during update, causing OOM on low-memory servers

Bug Description

hermes update restarts the gateway service as part of its update flow. On a low-memory server (1.6 GB RAM), this causes an Out-of-Memory kill because both the update process (git pull + pip install) and the gateway process (Python + Node/TUI) compete for memory simultaneously.

Environment

What Happened

1. Ran hermes update --yes from a terminal session
2. Update completed git pull + pip install successfully
3. Update attempted to restart hermes-gateway.service
4. Gateway restart triggered OOM — kernel killed the node process (TUI/Ink) inside the gateway cgroup:

oom-kill: task_memcg=hermes-gateway.service, pid=74547 (node)
Out of memory: Killed process 74547 (node) total-vm:12.9GB, anon-rss:976MB
hermes-gateway.service: Failed with result 'oom-kill'

1. The update command was interrupted (exit code 130 / SIGINT)
2. Gateway eventually restarted via systemd, but the update session was lost

A watchdog script had to be used as a workaround to ensure the gateway came back up.

Root Cause

The update flow runs pip install + git pull (memory-intensive) while the gateway is still running (~500 MB Python + ~1 GB Node/TUI), then restarts the gateway — which loads new modules while the old process is still draining. On a 1.6 GB server, the combined memory usage exceeds physical RAM and triggers OOM.

Proposed Fix

Option A (preferred): Stop the gateway before the heavy update steps (git pull, pip install), then start it after. This avoids memory overlap:

1. systemctl --user stop hermes-gateway
2. git pull + pip install
3. systemctl --user start hermes-gateway

Option B: Detect available memory before restarting. If memory is low, stop the gateway first, then restart after dependencies are updated.

Option C: At minimum, document this behavior and add a caveat for servers with < 2 GB RAM.

Impact

Users running Hermes on lightweight cloud instances (1-2 GB RAM) will experience gateway downtime after hermes update. Without a watchdog or manual intervention, the gateway may not recover.