Author: Lian Ran

Source: GeekPark

In recent weeks, an open-source project called Hermes Agent has become popular on X and GitHub.

From gaining over 22,000 stars in its first month of open source release at the end of February, to adding over 6,400 stars in a single day after the release of version v0.8.0 on April 8, Hermes Agent has amassed over 47,000 stars on GitHub in less than two months, and has consistently topped the global open source charts for several days.

What is a Hermes Agent?

In short, it's a personal AI agent that "self-improves": it has a built-in learning loop that can automatically extract skills from tasks, persistently remember user preferences, and accurately recall information across conversations, becoming more and more intuitive the more you use it; it supports 6 deployment methods including $5 VPS, Docker, and Serverless, and is compatible with 200+ large models. It can switch between Weibo, Telegram, Discord, and Slack platforms with a single command and can be installed and run with a single command.

Some say it's a replacement for OpenClaw, while others say it's even better than OpenClaw. In a very short time, it has not only garnered tens of thousands of stars, but has also made developers marvel that AI can "become more and more like a part of oneself the more you use it."

What exactly is a Hermes Agent?

Hermes Agent is a free, MIT-licensed, autonomous AI framework developed by Nous Research. Its core concept is a "self-evolving agent" that grows with use.

 Image source: GitHub

Unlike traditional agents, Hermes aims to be a long-term system that can continuously accumulate experience: it learns from completed tasks, retains memories across different sessions and platforms, and gradually forms a capability structure that belongs to the user.

 Image source: GitHub

Since its release in February 2026, it has garnered over 22,000 GitHub stars in just a few weeks, and currently boasts over 47,000 stars with hundreds of contributors. Community feedback suggests that this design addresses a long-standing need—developers are increasingly concerned with whether agents can "remember" and "become stronger."

In discussions within X and the developer community, what is particularly impressive is Hermes' significantly reduced reliance on cue words in complex tasks.

Some developers have mentioned when testing with gemma 26B or Hermes series models that even when given only a relatively vague instruction, such as "write a script to scrape data and generate visualizations," the Agent can complete the entire process from task breakdown to code generation.

During execution, it continuously adjusts its path based on execution feedback—including reading error messages, attempting to fix problems, and even forming reusable solutions through multiple attempts.

This experience does not mean that Hermes has stable "fully automated development capabilities", but it at least lets developers feel that the agent no longer relies entirely on precise prompts and can also advance complex tasks with ambiguous goals.

What's more noteworthy about Hermes Agent is actually its underlying architecture.

Architecturally, it and OpenClaw follow almost opposite paths: the former emphasizes the breadth of connectivity, while the latter focuses more on the depth of learning capabilities. Hermes' slogan—"the agent that grows with you"—points to a complete underlying design built around a "learning loop."

This closed loop consists of roughly three parts.

First, there's persistent memory. Hermes stores all historical conversations in a local database and reorganizes them through full-text search and model summarization. It can not only recall conversations from weeks ago but also establish connections between different tasks, gradually forming an understanding of how users work. This memory no longer relies on manual maintenance but is autonomously organized and updated by the agent, resembling a continuously evolving cognitive structure.

Secondly, there's the automatic generation and reuse of skills. When Hermes completes a complex task, it doesn't simply stop there; instead, it abstracts the entire process into a structured skill—including steps, key judgments, potential pitfalls, and verification methods. The next time it encounters a similar problem, it prioritizes drawing upon this existing experience rather than re-reasoning it. With increased usage, these skills are continuously refined and optimized, forming truly reusable capability assets.

Third, the rudiments of self-training capability. Hermes can generate a large number of tool call traces during operation and export this data for subsequent model fine-tuning. This means that it is not only "using the model" but also continuously producing training data that can feed back into the model. This capability already clearly has the characteristics of a research system, rather than a simple application-layer product.

This is why Hermes resembles an "experimental operating system" in its overall form. This tendency has become more pronounced in recent versions. For example, the introduction of multi-instance configuration allows developers to run multiple isolated agents in the same environment, each with its own independent memories, skills, and configurations. This allows Hermes to evolve further from a "personal assistant" into a reusable agent infrastructure.

For example, support for MCP allows it to expose its session and memory to IDE tools, enabling developers to directly retrieve and access them in Claude Desktop, Cursor, or VS Code. This design essentially breaks down the boundary between the "resident agent" and the "development environment."

In terms of security, Hermes appears relatively restrained. It uses mechanisms such as container isolation, read-only file systems, and pre-execution scanning to constrain potential risks. This design does not pursue extreme openness, but rather seeks a balance between "evolvability" and "controllability."

Overall, Hermes is not a project that excels in scale or ecosystem. Whether considering the number of GitHub stars or the maturity of the skills market, it is still in a relatively early stage.

What it truly deserves attention for is the direction it's betting on: how to make the agent stronger over time.

Similarities and differences between Hermes and Openclaw

When Hermes Agent shines on X, developers inevitably compare it with another phenomenal project that became a hit in the open-source community earlier this year—OpenClaw.

OpenClaw is also a local-first personal AI assistant framework. Both it and Hermes Agent attempt to address the privacy and control issues of traditional SaaS-based AI, but they diverge in their underlying philosophies.

The common ground between the two lies in "digital sovereignty".

Whether it's Hermes or lobsters, they all shared the same underlying genes when they first appeared:

• Local priority and privacy first: Data will not be uploaded to uncontrollable commercial cloud. All memories, code execution processes, and even file and directory level authorizations are kept on the user's local device or private environment as much as possible.
• Interaction based on messaging channels: They have all abandoned cumbersome web UIs and instead embraced instant messaging tools such as Telegram and WhatsApp, allowing AI to truly integrate into the daily communication chain of humans.
• 24/7 Automation: Supports scheduled tasks and can run silently in the background without requiring constant human monitoring.

This emphasis on "digital sovereignty" is essentially a choice at the infrastructure layer. Behind this, the differences between Hermes and OpenClaw are already beginning to emerge. Nous Research positions itself as a "decentralized AI research lab," not only developing agents but also advancing a decentralized training network called Psyche—attempting to use blockchain to coordinate globally idle GPUs for training large models.

Therefore, Hermes Agent is not only a local tool, but also the closest entry point to users in this whole set of "AI decentralized infrastructure".

The difference between the two lies in their different paths of capability development. If OpenClaw represents a more "deterministic" path, then Hermes is closer to an "evolutionary" system.

First, there's the difference in how skills are acquired (Human-authorized vs. Autonomous). The Lobster's capabilities are primarily determined by "human pre-defined rules." The Lobster prefers developers to write skills through explicit code or prompts. It's a perfect control plane; the user defines what it can do, and it executes it with high stability and determinism.

Hermes Agent's capabilities emerge through "experience emergence." After completing a complex task, it automatically abstracts methodologies, distills them into reusable skills, and continuously iterates and optimizes them in subsequent tasks. This means that its capabilities are not pre-defined but gradually expand during use.

Secondly, there are differences in memory mechanisms.

OpenClaw leans more towards explicit memory and retrieval mechanisms, essentially following the typical RAG approach—it knows "where the information is" and retrieves it when needed.

Hermes Agent employs a layered memory system. In addition to explicit memory, its core feature is the creation of a "model about you." Through cross-session interactions, it gradually understands the user's coding style, tolerance for errors, and preferred technology stack. It even periodically "nudges" itself to organize and solidify this knowledge.

The two also differ in their applicable scenarios.

If a user needs a highly secure, clearly defined task involving large volumes of data or financial transactions—tasks with extremely low tolerance for error—Lobster's access control (latest version) is more stringent and its behavior more predictable. However, if the user is engaged in exploratory programming, creative development, or handling complex projects with ambiguous processes requiring constant trial and error, the autonomy offered by Hermes will significantly reduce their workload.

However, the consensus in the communities on Reddit, YouTube, and X is not that Hermes replaced OpenClaw, but rather that they are complementary.

OpenClaw is responsible for "doing the work"—handling multi-channel interactions, team workflows, and complex ecosystem integration; while Hermes is responsible for "thinking"—focusing on persistent memory, automatic skill generation, and high-dimensional model reasoning.

A common setup is to run Hermes as an advanced planner on top of the OpenClaw tool. Simply running the `hermes claw migrate` command allows you to seamlessly migrate your existing OpenClaw skills, memories, and settings to Hermes with a single click.

"An Evolving Agent"

Hermes Agent is trying to transform "AI capabilities" from a one-time use to an asset that can be continuously accumulated.

An agent should not just be a temporary interface, but a long-term system—it is private, continuously running, and can accumulate capabilities during use, ultimately influencing the model itself.

Most mainstream products store their data, memories, and behavioral patterns on the platform side, but Hermes aims to keep these capabilities within the user's own system as much as possible.

This means that AI's capabilities are no longer just "called," but can be "owned." In the open-source community, the popularity of Hermes Agent largely stems from its successful implementation of a portion of this approach.

It is trying to establish a more complete chain: from task execution to skill accumulation, then to memory accumulation, and even further, to become part of the training data.

When an agent begins to possess this cyclical capability—solving problems on its own → recording experiences → reusing experiences → optimizing methods → and then feeding back into itself—it begins to approach a system capable of sustainable evolution.

Currently, this path is still in its early stages. Noise in memory, skill quality, and the stability of the training loop remain issues that require repeated refinement. Deployment barriers also persist, and there's still a considerable distance to go before it becomes seamlessly usable by ordinary users.

But the direction is already clear. Hermes has at least made one thing concrete: private AI is not just a form of use, but may be a form of asset that can continuously evolve.

If this path holds true, the way we evaluate an agent in the future may change: from looking at "what it can do now" to looking at "what it becomes over time." Whoever can accumulate more capabilities over time will have a higher ceiling.

Is the hottest agent project after "Lobster" suspected of being a scam?

Hermes' popularity is no accident; it's an open-source project with real technological output, but the project itself is not without controversy. And the biggest point of contention lies with the team itself.

Many of Nous Research's core members come from the Web3 field. Reports indicate that its CEO, Jeffrey Quesnelle, was previously the lead engineer at Eden Network, an Ethereum MEV infrastructure project.

The team's fundraising path also has obvious characteristics of the crypto industry - as of April 2026, Nous Research has completed two rounds of public financing, with a total financing amount of approximately $70 million. The investors are all leading institutions in the crypto field, and the fundraising path has distinct Web3 characteristics - it is priced in tokens rather than traditional equity, and the core funds are used for computing power reserves and team expansion.

Different sources of capital lead to different methodologies.

From the very beginning, Nous Research has been a Web3 native AI lab: it emphasizes decentralization in its governance structure, distributed training in its technical approach, and a "open source first + community driven" product strategy.

If we bring this background back to Hermes Agent, it's more like transplanting the methodology from the Web3 community to the infrastructure layer of AI Agent.

And it is precisely because of this that an issue that has been repeatedly discussed in the community has begun to emerge:

Does Hermes' "long-term operation + continuous accumulation" agent model also have the potential to become a kind of "Web3 cold start infrastructure"?

Currently, Nous Research remains in a "no token issued" state and has not explicitly announced any token distribution mechanism. However, in the broader ecosystem, some typical "anticipated behavior" can already be observed: for example, some crypto communities have begun discussing the anticipated airdrop around their project, and some third-party platforms are guiding users to participate in community interactions and task completion, using "potential rewards" as incentives.

At the same time, unofficial tokens named "NOUS" have also appeared on the blockchain. These assets are not directly related to the project itself, but they are often amplified and misinterpreted when market sentiment fluctuates.

These phenomena do not directly indicate the direction of the project, but they at least show one thing: the market is already using the logic of Web3 to "understand" the project in advance.

Structurally, Hermes Agent does indeed possess some characteristics common in the crypto world: it runs locally on the user's device and is online for extended periods; it continuously generates behavioral data and interaction patterns; and it constantly accumulates "contributions" during use.

In the traditional software context, these are just part of the "product experience"; but in the Web3 context, such behavior is often regarded as a "measurable engagement".

This puts Hermes in a relatively delicate position: on the one hand, it is a real, usable, and rapidly iterating open-source agent framework; on the other hand, its technical path and community structure naturally have the potential to extend to a "tokenized incentive system".

For developers, the value of Hermes Agent still primarily stems from its capabilities as an agent system. However, for ordinary users, this means a more realistic criterion: any transaction, investment, or promise directly linked to the "NOUS token" requires considerable caution—especially if there are already on-chain assets with the same name or high-yield promotions.