Note: This is a re-post of an article I published on my LinkedIn.

Innovation doesn’t wait for you

Every day, my feed filled with new OpenClaw clone, new frameworks, so many “this changes everything” posts. But, I was already OpenClaw on a VM and it was working fine. It was powerful and I made my customizations. So, I ignored the noise and resisted change.

Then Nous Research dropped Hermes Agent v0.6 with multi-agent support. That was the nudge I needed. I had previously tried Paperclip.ai for multi-agent orchestration and it just didn’t work for me.

So many claws

OpenClaw started this wave but the ecosystem is exploding. Some notable clones:

• IronClaw → Rust-based, more enterprise-ready
• ZeroClaw → minimal, faster, simpler
• NemoClaw → Nvidia’s security-focused version
• And many more: PicoClaw, MetaClaw, CoPaw, NanoClaw, NullClaw…

I’m sure there are many more. Development on OpenClaw moves fast, new major features are being released on a weekly basis. That’s something that typically takes months. We’re talking about things like agent capabilities, toolsets, critical security patches and workflow features like /tasks that change how you work. The question remains: do any of these new “claws” substantially improve things for my main issue? The main issue being memory.

The real problem: memory

I was running 2 agents:

• Agent 1 helps me run my home systems (Proxmox, HomeAssistant, Security cameras), social media, and research.

• Agent 2 is more personal, calendars, emails, reminders, birthday tracking and gift ideas.

It worked… until it didn’t. My main agent started getting confused. Like a “jack of all trades” employee with too many responsibilities and not enough specialization. To fix it, I added:

• lossless-claw → store everything
• ByteRover → find everything

It helped. But it still wasn’t great. Why? Because I was doing the work the agent should be doing. Things like telling it to remember, verifying that it could retrieve the facts, and managing plugins. At some point, you realize that the more plugins you add the more you become the orchestrator.

Hermes fixes memory at the system level

Here’s a comparison:

• OpenClaw: memory is external (bolt on), plugin-dependent, and fragmented.
• Hermes: memory is built-in, it’s selective and contextual and most importantly self-maintaining.

Let’s break it down:

Feature	OpenClaw	Hermes Agent
Persistence (storing facts)	needs lossless-claw or similar plugin	built-in
Retrieval (finding context)	needs ByteRover or other plugin	built-in
Synthesis (making memory usable)	not really there yet	core feature
Autonomous updates (deciding what matter)	manual	automatic

OpenClaw tries to remember everything. Hermes tries to remember what matters. That means less babysitting, better decisions and much cleaner execution.

Skills are the real unlock

Better memory was great by itself but then I noticed something wild! Autonomous skills.

Hermes started creating skills… on its own. No prompt, no instruction. For example, I asked it to fix a broken local site. It SSH’d in, fixed it, then created a reusable skill. I didn’t ask it to do that. Now it knows where the files are (/opt), the stack (NextJS, Tailwind) and how to fix it again in the future.

That’s not just automation. That’s self-improvement.

As of Hermes 0.6, it can spin up subagents on its own, run tasks in parallel, and share memory and skills. This is a shift from tools to teams. Separate agents means better outcomes, less confusion, just like real specialized employees.

What held me back

I hesitated because OpenClaw has a huge ecosystem, tons of plugins, and a massive community. But it didn’t matter because both Hermes and OpenClaw can use Clawhub (skills directory) and support shared skill ecosystems. So the gap isn’t as big as it looks.

We already see enterprises adopting this paradigm. Microsoft moved here. Omar Shahine recently talked about bringing OpenClaw-style agents into M365. When Microsoft moves, it’s not a trend, it’s a signal.

If you’re technical, spin it up with Docker in minutes. If you’re not, use a hosted option. OpenClaw has many. Hermes? Still early. I only found one viable hosted route (Hostinger). Later, Nous Portal announced their own hosted options (this is where I would go).

Final thoughts

We’re not just watching better AI tools emerge. We’re watching a shift from assistants that respond to agents that do, remember, learn, and delegate. We’re getting closer to Iron Man’s Jarvis. OpenClaw showed what was possible. Hermes is starting to show what’s sustainable. The real question now isn’t which one wins. It’s how long before we stop doing this work ourselves.