Since the start of 2026, the AI Agent landscape has continued to heat up. Just as OpenClaw maintained its lead as the top open-source Agent with 354,000 stars, a new project named Hermes Agent achieved a remarkable overtake at breakneck speed.

Topping the Open-Source Rankings

In late February 2026, renowned open-source large model institution Nous Research officially launched this product. With its core positioning of "growing together with users," it secured 22,000 stars in its first month. After the release of the v0.8.0 version in April, it gained over 6,400 stars in a single day, amassing more than 64,000 GitHub stars within six weeks. By mid-April, the total star count had surpassed 70,000, with an average weekly increase of approximately 9,500 stars—more than triple the growth rate of OpenClaw during the same period—dominating GitHub's global trending list for multiple consecutive days.

On the OpenRouter platform, Hermes Agent achieved a 367% explosive growth within a week, with usage reaching 971 billion tokens, sparking heated discussions about "replacing OpenClaw."

This surge precisely coincided with an industry window of opportunity. In early April, OpenClaw faced a dual crisis: Anthropic adjusted its policies, and Claude's subscription quotas no longer covered third-party frameworks.

Simultaneously, multiple high-risk CVE vulnerabilities (CVSS score 9.9) were exposed in OpenClaw. Nous Research swiftly launched the hermes claw migrate one-click migration tool, precisely meeting the migration demand.

Industrial capital quickly entered the scene: Xiaomi's large model, Xiaomi MiMo, achieved deep integration and offered limited-time free calls. MiniMax, Zhipu GLM, and others successively formed strategic partnerships, shedding the label of a "geek toy" and propelling its transformation into mainstream AI infrastructure.

In terms of deployment, Hermes Agent supports Linux, macOS, WSL2, and Android Termux, with installation completed via a single curl command. It can run 24/7 on a VPS costing as little as $5 per month.

It features a built-in unified messaging gateway, integrating with 15+ platforms including Telegram, Discord, Feishu, and WeCom, with fully interoperable memory and skill data, completely resolve (thoroughly resolving) OpenClaw's pain point of "fragmented memory across platforms."

Four-Tier Hierarchical Memory Architecture

The success of Hermes Agent is fundamentally a victory of Nous Research's technical accumulation and architectural innovation. Founded in 2023, the team of approximately 20 members hails primarily from StabilityAI, with cumulative funding of around $70 million, including a $50 million Series A round led by Paradigm in 2025.

Its previously launched Hermes and Nomos series models have accumulated over 50 million downloads, building deep expertise in tool invocation and long-term planning. The "model trainer turned Agent developer" DNA gives it a better understanding of large models' capability boundaries.

Unlike traditional "use-and-discard" Agents, Hermes constructs a three-tier closed-loop system of "memory-skills-training data." In memory system design, Hermes Agent takes a counterintuitive approach, eschewing full-volume storage and instead adopting a CPU cache-inspired hierarchical strategy to create a four-tier memory architecture.

The first layer, L1 core memory, stored in the MEMORY.md file, is strictly limited to approximately 800 tokens. Each session starts by freezing it as a snapshot injected into the system prompt, forcibly retaining the highest-value project context.

The second layer, L2 user persona, stored in the USER.md file, contains about 500 tokens, recording the user's communication style, technology stack preferences, and work habits.

The third layer, L3 long-term memory, employs SQLite FTS5 full-text retrieval technology. All cross-session dialogue histories are stored here, enabling rapid retrieval of tens of thousands of entries within 10ms. LLM summaries are recalled on demand to avoid irrelevant information interference.

The fourth layer, L4 skill library, stores procedural experience. By default, only skill name indices are loaded; full text is loaded only upon matching, preventing context explosion even with 600+ accumulated skills.

This design is compatible with Anthropic's prompt caching mechanism and reduces token costs.

Positive Feedback Loop

Hermes Agent achieves true autonomous evolution. After each task, it automatically executes a four-step closed loop: "execution-evaluation-abstraction-optimization." The Router matches skills, the Executor carries them out, and the Evaluator assigns scores (1-10).

If the score is ≥7, the task involves ≥3 steps, and no existing skills are reused, the Skill Extractor automatically refines the successful path into a structured Markdown skill document (including step sequences, pitfalls, and validation criteria), adhering to the agentskills.io open standard for cross-project migration and community installation.

Skills are updated through "patch-style modifications," only revising problematic segments. User testing shows that after a month of continuous use, tool invocations for similar tasks drop from over 20 to 8-10, improving efficiency by over 60%.

Integrating the Atropos reinforcement learning framework, it converts tool invocation trajectories into training data to feed back into the model, forming a positive feedback loop of "memory → skills → training → model → Agent." This data reflow (reflow) capability creates a barrier difficult for pure application or model layers to replicate.

Multi-Model Compatibility: Supports 12+ providers including Nous Portal, OpenRouter, Anthropic, OpenAI, Xiaomi MiMo, and Zhipu GLM, with flexible cost control via main + auxiliary models. Switching commands are completed with one click, and memory/skill data is saved locally, with zero migration cost. Paired with 47 built-in tools like the Camofox anti-detection browser and Tirith security sandbox, it balances functionality and security.

Polarized Evolution

According to Grand View Research, the global AI Agent market size was approximately $7.63 billion in 2025, expected to surpass $10.91 billion in 2026 (CAGR 49.6%) and reach $182.9 billion by 2033.

Hermes represents the self-evolutionary route, marking the sector's shift from "tool invocation" breadth competition to "capability evolution" depth competition.

Traditional Agent frameworks (OpenClaw, AutoGPT) solve "orchestration" problems but reset context after each session, failing to accumulate experience. Memory relies on manual maintenance, skills require manual coding, and cloud hosting locks data to the platform.

Hermes is the first to engineer "self-evolution" in the open-source space, enabling AI to understand users better over time.

The open-source Agent landscape is polarizing: one end features the "enterprise infrastructure" route (e.g., OpenClaw, Claude Code) with strong permission controls and audit logs; the other end is the "self-evolving digital partner" route (e.g., Hermes) emphasizing personalization and continuous learning.

These are not zero-sum; many users run Hermes as an advanced planner on OpenClaw's tool ecosystem, forming a "Hermes planning + OpenClaw execution" model.

By late March 2026, DingTalk, Feishu, and WeCom open-sourced CLI tools, with enterprise applications officially shifting toward "serving AI agents." Hermes natively supports Feishu and WeCom, gaining first-mover advantages in Chinese-language scenarios.

It maintains zero CVE vulnerabilities in security, with built-in Tirith pre-execution scanning, path traversal protection, and MCP OAuth 2.1, strongly appealing to enterprise clients.

The business model adopts "open-source acquisition + cloud service monetization": the core framework is open-sourced under the MIT license, with matching (supporting) Nous Portal subscription services (zero-configuration access to 400+ models). It builds ecological barriers via agentskills.io, following successful precedents like MongoDB and Elastic.

Core Trends

We must also remain soberly aware that Hermes Agent and the entire self-evolving AI sector still face multiple challenges.

In terms of technical maturity, Hermes has only been released for two months (v0.8.0), with long-term stability yet to be verified. Ecologically, OpenClaw boasts 13,000+ community skills and 24+ platform adaptations, requiring long-term construction by Hermes.

Commercially, the profitability timeline remains unclear, and Nous's encryption/Web3 background may raise concerns among some enterprise clients.

Model dependency requires underlying models to support at least 64K context for self-evolution effects, with high-end APIs being costly for individual users. Security-wise, 24/7 operation risks data leaks; despite 50+ security fixes, sandboxing and privacy protection still need optimization.

Long-term, the transformation of AI Agents from "one-time tools" to "long-term digital partners" is irreversible. Hermes has already natively supported WeChat Callback mode in v0.9.0, continuously strengthening enterprise-grade capabilities.

When Agents begin accumulating skills, generating training data, and feeding back into models independently, true "self-evolving AI systems" will no longer be distant.