Agent Infrastructure
OpenClaw had the infrastructure. Hermes had the memory. Memory won.
As of May 10, 2026, Hermes Agent — a six-week-old open-source agent from Nous Research — is processing 224 billion tokens per day on OpenRouter. That is more than OpenClaw, more than every proprietary chat app, more than every other application on the platform. It is the most-used AI application on OpenRouter, period. And the reason it won is not what anyone in February would have predicted.
The numbers that ended the debate
On May 6, Nous Research announced that Hermes Agent had hit #1 on OpenRouter's global rankings with 271 billion tokens processed. Four days later, OpenRouter's public rankings showed Hermes at 224 billion daily tokens versus OpenClaw's 186 billion. The gap is widening.
This is not a niche leaderboard for CLI tools or agent frameworks. OpenRouter is a unified API gateway routing requests to 200+ models across every major provider. Its rankings track all applications — chat UIs, productivity tools, IDEs, enterprise platforms. Hermes is not just the #1 agent. It is the #1 app.
For context: Hermes Agent launched on February 25, 2026 with a quiet tweet. Six weeks later it had 57,200 GitHub stars, 80+ ecosystem projects, 8 external memory providers, 9 multi-agent orchestration frameworks, and a growth rate of roughly 5,000 ecosystem stars per week. OpenClaw's founder joined OpenAI in February, and the project moved to an independent open-source foundation. The guard changed fast.
Two architectures, one winner
The rivalry between Hermes and OpenClaw is not a brand war. It is a referendum on what an AI agent should be.
OpenClaw is built around a central WebSocket Gateway — a persistent routing layer connecting 50+ messaging channels (Telegram, Discord, Slack, WhatsApp, Signal, SMS) to an agent runtime. Its design optimizes for reach: how many surfaces the agent can operate across simultaneously. The thesis: be everywhere, and value follows ubiquity.
Hermes Agent takes the opposite approach. Under an MIT license, it centers on a "do, learn, improve" execution loop. After completing a complex task — debugging a microservice, refactoring a codebase, researching a competitor — it synthesizes that experience into a permanent, reusable skill document. Next time, it does not start from zero. It loads the relevant skill and gets to work. The thesis: get smarter with every interaction, and value follows intelligence.
The market just voted. OpenClaw's reach-first architecture produced a tool that connects to everything but remembers nothing across sessions. Hermes's learning-first architecture produced an agent that gets more capable the longer you use it. That second thing turned out to be the one users actually wanted.
The memory advantage is not subtle
Hermes maintains three layers of memory. At the top, a bounded MEMORY.md and USER.md file always live in the system prompt — the agent never forgets who you are or what matters. Below that, every session is stored in SQLite with FTS5 full-text search, so the agent can recall any past conversation. Above that, eight pluggable external memory providers (Honcho, Mem0, Hindsight, Supermemory, and others) handle long-term semantic memory and graph-based recall.
This is the architectural decision that won. A chat that remembers your last ten projects, your coding style, your deployment preferences, and which bug you fixed last Tuesday is fundamentally more useful than one that asks "What can I help you with today?" every Monday morning.
OpenClaw can do many things Hermes cannot — particularly on the multi-platform messaging front. But the thing users most wanted was an agent that accumulated knowledge instead of resetting. Hermes delivered that, and the token volumes reflect it.
What this means for the agent ecosystem
The leaderboard swap has implications beyond one project overtaking another.
Open-source is winning at scale. Hermes is MIT-licensed and self-hosted. Users run it on their own VPS, laptop, or home server. They choose their own models (20+ providers supported). They control their own data. This is not a proprietary SaaS play — it is infrastructure you own, and the market is choosing it over polished hosted alternatives.
The skill ecosystem is a moat. Hermes ships with 40+ bundled skills covering MLOps, GitHub workflows, diagramming, note-taking, and more. The community has added 17 skill libraries, including a 754-skill cybersecurity collection from Anthropic mapped to MITRE ATT&CK. Each new skill makes every Hermes instance more capable — and makes switching to another agent framework more expensive.
Memory architecture is now a competitive requirement. Every AI agent framework shipping in the second half of 2026 will need to answer the question: "What does your agent remember, and for how long?" Hermes raised the bar from "session history" to "persistent, searchable, self-improving knowledge." Frameworks that ship without an answer will lose.
The counterintuitive lesson
If you had asked anyone in February 2026 which AI agent would dominate OpenRouter by May, the smart money was on OpenClaw. It had the GitHub stars, the multi-platform reach, the brand recognition. Hermes had a good idea and a research lab behind it.
What nobody predicted: the winner would be determined not by how many platforms the agent reached, but by how much it remembered. OpenClaw optimized for breadth. Hermes optimized for depth over time. Depth over time won.
The lesson for builders: in the agent space, persistence beats ubiquity. Users will tolerate fewer platforms if the agent on their terminal and Discord actually learns from every interaction. They will not tolerate an agent on every platform that treats every conversation like the first one.
The practical takeaway
If you are building with agents, the Hermes-OpenClaw dynamic teaches one thing: invest in the architecture that compounds. A skill system that improves with use, a memory layer that deepens across sessions, a user model that actually learns preferences — these are not nice-to-have features. They are the difference between an agent people try once and an agent they run in production.
Try Hermes Agent yourself: github.com/NousResearch/hermes-agent. The quickstart gets you from clone to a running agent in under five minutes. Start with one platform — terminal or Discord — and give it a real task. Let it write a skill. Then give it a similar task tomorrow and watch what happens.
That second interaction is where Hermes wins. And now the whole industry knows it.