Hermes Agent Finally Fixes Everything in Its Make-Up Update
Hermes v0.22.2 update: what changed in this release, explained in plain language.
Version Info
| Item | Details |
|---|---|
| Version | v0.22.2 |
| Release Date | October 15, 2026 |
| Release Theme | Re-delivering what v0.22.1 was supposed to deliver |
| One-liner | v0.22.1 was packaged with the wrong code, which pulled in unreleased gateway features and left the Microsoft Store version stuck on “Connecting.” v0.22.2 is what v0.22.1 should have been. |
Known minor issue: Sessions that ran under the 0.22.1 gateway can’t be deleted in 0.22.2 for now.
Plain-English Breakdown
After a reboot, the gateway can start itself again
Think of the gateway as Hermes’s “switchboard” — it routes incoming messages to the agent. There used to be a sneaky bug: after a computer restart, Linux would sometimes assign an old process ID (PID) to some unrelated program. The gateway would see that ID and mistakenly think “the switchboard is already running,” so gateways managed by supervisors like systemd, launchd, or Docker would stubbornly refuse to start.
Now Hermes remembers which boot cycle it belongs to. When it really can’t tell, it checks a “door lock” — only a genuinely alive gateway holds the key. No more mistaken identity, and the switchboard gets back to work after every reboot.
Running local models on a laptop: long conversations no longer die mid-stream
Lots of people like running models locally — it saves money and keeps things private — but laptops go to sleep and memory gets tight. Here’s what changed:
- Local requests now always use streaming, and the status bar shows the model’s progress through the prompt, so you can see it’s actually working instead of just waiting;
- The stuck timer now skips over time the machine spent asleep, so closing your lid for two hours won’t trigger a “timed out” failure;
- After the local service recovers, it automatically retries once instead of blacklisting all subsequent messages after a single failure;
- Conversation compression no longer times out just because the server is still reading a long conversation;
- On Windows machines with unified memory, model startup and prompt caching are allocated based on available memory, so it’s harder to crush the machine.
A few small conveniences thrown in
The model picker now has a “curated list” at the top, so you don’t have to scroll through a giant list; nemotron has been added to NVIDIA NIM’s default models with sensible default parameters. Also, Microsoft Store packaging has been split into its own windows-store channel, so future Store updates will have a cleaner cadence.
Various gateway message-handling fixes
This release also tackles a batch of gateway “relay” issues: internal wakeups no longer incorrectly attach replies to the triggerer; rejected process-completion notifications back off for a few minutes before sending a plain reminder instead of spamming repeatedly; completion wakeups that are just missing a message_id are now received correctly; and after a session-following switch, if the original profile loses focus, it properly drops its “coding posture.”
Summary
For regular users, the most noticeable thing about this update is: Hermes starts up properly after a reboot, long conversations with local models are more stable and less likely to break, and model selection is less of a hassle. It’s essentially a “paying off the debt” release — no flashy new features, but it fills in the holes v0.22.1 left behind. If you’ve been plagued by “Connecting” or the gateway refusing to start after a reboot, this version is worth upgrading to.
📎 Official release notes: https://github.com/NousResearch/hermes-agent/releases