Hermes Agent learns from its own mistakes and gets 40% faster — but don't call it self-improvement

No AI agent makes itself smarter. That's a lie the industry sells you wrapped in buzzwords. Hermes Agent from Nous Research is something far more interesting — and far more honest: it's an agent that gets faster by remembering what worked last time. That's not magic. That's architecture.

From zero to 60,000 stars in four months

Nous Research released Hermes Agent on GitHub on February 25, 2026, under the MIT license. The project grew explosively — from 42,000 stars in April to a conservative estimate of 60,000+ today, according to data from cognio.so, nxcode.io, and innobu.com. The codebase is 93.8% Python and runs locally, in the cloud, or on European infrastructure.

Two milestones have defined the project so far:

> v0.8.0 (April 8, 2026): First production release with 209 merged PRs, Google AI Studio integration, MCP OAuth 2.1 with PKCE, and the first version of GEPA.

> v0.10.0 (April 16, 2026): 118 built-in skills, three-layer memory architecture, six messaging integrations, and a closed learning loop.

FACT BOX: What is Hermes Agent?

GEPA: The algorithm nobody talks about honestly

Here's the core that everyone gets wrong. GEPA — Genetic-Pareto Evolution — was accepted as an Oral Presentation at ICLR 2026, meaning peer-reviewed AI researchers deemed it worth a spotlight (innobu.com, petronellatech.com). That's not marketing. That's academic validation.

But what does it actually do?

GEPA uses a language model to read through complete execution traces — every step the agent took, every error message, every profiling data point, every reasoning chain. Based on this, it proposes targeted prompt improvements. Not via a simple reward function that collapses everything into a single scalar, as traditional reinforcement learning does. Instead, it preserves multiple objectives simultaneously: speed, reliability, resource use.

The result: 40% faster on repeated tasks in the same domain once the agent has built 20 or more self-generated skills (Nous Research).

> "40% faster is not the same as 40% smarter. GEPA optimizes for efficiency, not intelligence. The same underlying model produces the answers — it just finds them faster."

> — innobu.com, April 2026

This is critical to understand. Hermes Agent with GEPA is like an experienced surgeon versus a resident: both have the same training, but the experienced one knows which tools to reach for without searching. The underlying expertise is identical.

The four-step learning loop — how it actually works

1. Execute

The agent completes tasks using 40+ built-in tools — from code search and file manipulation to API calls and web browsing.

2. Evaluate

Results are assessed via explicit feedback (you say "that was good" or "that was wrong") and implicit acceptance signals (task completion without correction, successful code execution).

3. Abstract

Successful patterns are converted into reusable skill documents stored as Markdown files in ~/.hermes/skills/. This is the agent's long-term memory.

4. Refine

When a similar task appears again, the agent retrieves relevant skills and improves them based on new experience.

KEY FIGURES

118

Built-in skills in v0.10.0

~10ms

Search time via FTS5 full-text search across 10,000+ skills

7 days

The Curator's automatic cleanup cycle

200+

Supported models via OpenRouter

Three-layer memory architecture: More than a context window

Most AI agents live and die within their context window. Hermes Agent is designed differently with three distinct layers:

Short-term memory: Standard session context for what you're working on right now.

Persistent long-term memory: A SQLite database with FTS5 full-text search that can find relevant experiences in around 10 milliseconds even across 10,000+ stored skills. That's faster than most humans can turn their heads.

Procedural skill memory: LLM-generated summaries of successful workflows, stored in structured form and accessible for future tasks.

The Curator cleans up while you sleep

From version 0.12, Nous Research introduced The Curator — a background process that reviews the entire skill library on a rolling 7-day cycle. It consolidates overlapping skills that have been duplicated, archives entries that are no longer relevant, and improves the descriptions of active skills based on recent use.

This matters because uncontrolled skill accumulation is a real problem. Without maintenance, the library would grow unmanageable and search quality would degrade. The Curator is the solution nobody talks about but everyone needs.

Model agnosticism: Run what you want

Hermes Agent doesn't lock you into one vendor. Via OpenRouter, over 200 models are supported. You can use:

• Commercial: OpenAI, Anthropic Claude, Google Gemini
• Chinese alternatives: DeepSeek, Qwen, Xiaomi MiMo, z.ai GLM, Kimi Moonshot, MiniMax
• Local models: Via Ollama on your own hardware
• Nous's own models: Via Nous Portal

And you can switch models mid-session without restarting. In practice, that means you can start an analysis with a fast budget model, switch to Claude or GPT for complex reasoning steps, and return to a local model for formatting — all in one workflow.

EU AI Act: The clock is ticking

The deadline is approaching. August 2, 2026, general obligations under the EU AI Act take effect. For European organizations considering Hermes Agent, the situation looks like this:

Arguments in favor: Hermes Agent can be self-hosted on European infrastructure, has zero telemetry, and maps naturally to GDPR's principle of local data processing. The project is open source, so you can inspect everything.

What's required: Audit logging, transparency documentation, and review of the skill library (skill vetting) for AI Act compliance. The Microsoft Agent Governance Toolkit (released April 2, 2026, MIT license) provides concrete compliance patterns that can be applied directly.

Bottom line: Hermes Agent isn't pre-packaged for EU AI Act compliance out of the box. It requires work. But the foundation is better than most proprietary alternatives.

TIMELINE: Hermes Agent from zero to production

February 25, 2026 — Launched on GitHub under MIT license

April 8, 2026 — v0.8.0: First production release, GEPA introduced, Google AI Studio integration

April 16, 2026 — v0.10.0: 118 skills, three-layer memory, closed learning loop

v0.12+ — The Curator: automatic skill library maintenance

August 2, 2026 — EU AI Act general obligations take effect

Limitations you cannot ignore

This is where many people stumble. Because Hermes Agent is impressive on paper, it's easy to forget this is a v0.x project:

• Documentation is incomplete. You will hit walls where good answers don't exist.
• API stability is not guaranteed between minor releases. Code that works in v0.10 may require changes in v0.12.
• Not ready for critical production — ERP systems, customer data, or mission-critical workflows should stay away for now.
• Young community means slower response times on GitHub Issues and fewer Stack Overflow answers.

For pilot projects in developer productivity, research assistance, and routine automation, the risk profile is reasonable. For everything else: wait.

BOTTOM LINE

Hermes Agent is the most promising open AI agent project of 2026, and the GEPA mechanism — peer-reviewed and accepted as Oral at ICLR — gives it academic credibility most competitors lack. The 40% speed gain on repeated tasks is real and measurable.

But it is not self-improvement in the science fiction sense. It doesn't get smarter. It gets more efficient. That's a crucial distinction.

For teams looking to experiment with AI agents at a cost of €5–15 per month, with zero telemetry and full data control: Hermes Agent is worth a serious pilot project. For those who think they're buying an agent that gradually surpasses human intelligence: go home and sleep on it.

Verified against 4 open primary sources (GitHub, innobu.com, petronellatech.com, cognio.so) and 2 independent analyses (nxcode.io, Nous Research official documentation).