OpenClaw vs. Hermes Agent: The race to build AI assistants that never forget

Host A: Welcome back to DevTools Radio, I'm here with my co-host, and today we're diving into something that I think every developer listening has felt deeply in their bones — that moment when you close an AI coding session and realize everything you just taught it is just... gone.

Host B: Oh, the pain is real. You spend hours explaining your legacy database schema, your bizarre naming conventions, the deployment quirks that nobody ever bothered to document — and then poof. Next session, you're starting from scratch like it never happened.

Host A: Exactly. And two open-source projects are now going head to head trying to solve that problem — OpenClaw and Hermes Agent. They're both pushing toward what you might call a persistent agent runtime, but they're taking wildly different approaches to get there.

Host B: So walk me through OpenClaw first, because the growth numbers on this one are kind of staggering.

Host A: Right, so OpenClaw started as a weekend project by Austrian developer Peter Steinberger in late 2025 — originally called Clawdbot — and it exploded to over 345,000 GitHub stars by early April 2026. The core idea is that your AI agent lives permanently on your infrastructure, connects to WhatsApp, Telegram, Slack, Discord, over 50 integrations, and it's running while you sleep.

Host B: So it's less of a tool you open and more of a presence that just... exists alongside you. That's a genuinely different mental model for most developers.

Host A: The analogy I keep coming back to is Android for AI agents — massive ecosystem, tons of third-party skills on their ClawHub marketplace, model-agnostic so you're not locked to one provider. But, and this is a big but, it carries Android's early baggage too.

Host B: Meaning the security story is, let's say, a work in progress?

Host A: That's being generous. A security audit found 341 malicious skills out of roughly 2,857 on the marketplace. There's a CVE with an 8.8 severity score around unsafe WebSocket behavior. Cisco called personal AI agents like OpenClaw "a security nightmare." Microsoft advised against running it on standard enterprise workstations. Publishing a skill initially required nothing more than a week-old GitHub account.

Host B: Okay so that's the breadth play — huge ecosystem, huge reach, real security headaches. Where does Hermes Agent fit in?

Host A: Hermes Agent comes from Nous Research, launched February 2026, and it's sitting at around 22,000 stars — so a fraction of OpenClaw's community. But the architecture underneath is genuinely fascinating. Their whole philosophy is depth over breadth. The tagline is literally "the agent that grows with you."

Host B: And how does that growth actually work mechanically? Because a lot of tools claim memory features that turn out to be pretty shallow.

Host A: Three things make it different. First, it uses FTS5 full-text search over all past sessions stored in SQLite with LLM-powered summarization — so it can actually recall a conversation from three weeks ago and build on it. Second, after completing complex tasks, it autonomously writes structured skill documents recording the procedures and pitfalls it discovered. And third — this is the wild one — it integrates with a reinforcement learning framework to actually train on its own behavior over time.

Host B: So it's not just remembering what you told it, it's genuinely getting better at the job the longer it runs. That's a pretty different value proposition than anything bolted onto a session-based tool.

Host A: Exactly. And to be fair to tools like Claude Code and Cursor, they're adding persistence features too — auto-memory, workspace context — but OpenClaw and Hermes Agent are designed from the ground up for this. The ambition is architecturally different.

Host B: So where does that leave developers who are listening right now trying to figure out which direction to bet on?

Host A: Honestly, I'd say OpenClaw wins if you want something with a massive community and you need cross-channel reach right now — just go in with eyes open about the security risks and treat the skill marketplace like early npm, with appropriate skepticism. Hermes Agent is the one to watch if you care more about a learning loop that actually compounds over time. Smaller community today, but the foundation is more sophisticated.

Host B: The stateless container versus the persistent volume, to use the infrastructure analogy from earlier. One's convenient, one's invested in you long-term.

Host A: That's exactly it. And whichever wins, the session-based AI assistant that forgets everything at close — its days are numbered. Alright, that's going to do it for today's deep dive on DevTools Radio.

Host B: Thanks for listening, everyone. If you're experimenting with either of these projects, we'd genuinely love to hear how it's going — find us wherever you get your podcasts. See you next time.