Something fundamental shifted in February 2026. Nous Research released Hermes Agent — an open-source autonomous AI agent — and within seven weeks it had crossed 95,000 GitHub stars. As of May 2026 it sits at over 105,000, making it the fastest-growing open-source agent project of the year. The reason is not the star count. The reason is that Hermes solves a problem every developer who has worked with AI agents will immediately recognise.
AI agents forget everything the moment the session ends. You explain your codebase on Monday, and by Friday the agent has no memory of it. You solve a complex deployment problem with careful prompting, and next week you solve it again from scratch. This is not a fringe complaint. It is the core limitation of every agent framework shipped between 2023 and 2025. Most of them tried to fix it with vector databases and called it “long-term memory.” It was not.
What Hermes Agent Actually Is
Hermes Agent is an open-source autonomous AI agent built by Nous Research, released on 25 February 2026 under the MIT licence. It is not a coding copilot tied to an IDE, not a chatbot wrapper around a single API, and not a hosted SaaS product. It runs on your own server — a $5 VPS, a local machine, a GPU cluster — stores all data locally, and sends no telemetry to any third party.
The core distinction is the learning loop. When Hermes solves a problem that requires five or more tool calls, it automatically writes a skill document: a structured, searchable record of what worked, what the steps were, and what tools were used. The next time a similar problem appears, the agent loads that skill, applies the prior procedure, and refines it if something has changed. That is a fundamentally different architecture from session-based agents. It is closer to how a competent human employee works — building institutional knowledge over time — than to a stateless API call.
The Three Core Systems
1. Persistent Memory
Hermes stores all memories, skills, and conversation history in a local SQLite database with FTS5 full-text search. When you start a session, the agent searches its own past sessions for relevant context using LLM-powered summarisation. It remembers project preferences from last week and can recall task details from three months ago. The memory does not decay. Nothing is stored in a cloud service. You own everything.
2. Autonomous Skill Creation
After any complex task involving five or more tool calls, Hermes automatically creates a skill document. Skills are structured, human-readable, and compatible with the agentskills.io open standard, which means they are portable across frameworks. The Autonomous Curator — a background process — regularly reviews the skill library, consolidates overlapping skills, archives stale entries, and writes per-run reports. By v0.13.0, the Curator had become one of the most refined parts of the system.
3. Self-Improvement Loop
Skills are not static. Each time a skill is used, Hermes patches it if the steps were outdated, incomplete, or wrong. The companion project — Hermes Agent Self-Evolution, an ICLR 2026 Oral paper from MIT — applies DSPy and GEPA to optimise skills, prompts, and the agent’s own code against benchmarks. The question of whether this produces genuine compounding improvement or better UX is still live research, but the mechanism is real and verifiable by reading a file on disk.
What It Supports
As of v0.14.0 (released 16 May 2026), Hermes Agent supports:
- 40+ built-in tools across file operations, web search, code execution, image generation, and browser automation
- 14+ messaging platforms including Telegram, Slack, Discord, Feishu, and WhatsApp — with conversations continuous across platforms
- Model Context Protocol (MCP) in both client and server modes — connect Claude Desktop, Cursor, VS Code, and other MCP clients
- Profiles system — multiple isolated Hermes instances from a single installation, each with separate memories, skills, credentials, and configurations
- Honcho integration for dialectic user modelling that builds a persistent model of who you are across sessions
- Pluggable memory backends (v0.7.0) for extended storage options
- Linux, macOS, and WSL2 — single curl install command, no prerequisites
Why It Matters in 2026
The Stanford HAI AI Index 2026 reported that on OSWorld, agent accuracy rose from roughly 12 per cent to 66.3 per cent between 2024 and 2025 — within six percentage points of human performance. Simultaneously, frontier model capability is converging: four major labs were clustered within 25 Elo points in Arena ratings as of March 2026. When raw model intelligence compresses, differentiation moves to the outer layer: memory systems, tool execution, reliability, and cost control. Hermes is built precisely at that outer layer.
The open-source positioning is also deliberate. Every other serious agent in the 2026 landscape — Cursor, Claude Code, OpenClaw — involves some degree of cloud dependency or commercial licensing. Hermes is MIT-licensed, local-first, and designed for operators who want full control over what their agent stores, learns, and does.
Who Should Use It
Hermes Agent is worth serious attention if you:
- Use Claude Code heavily and want it to remember your projects across sessions
- Do content work, research, or data analysis that benefits from persistent context
- Want a private AI assistant running on your own hardware with no cloud exposure
- Are building agent infrastructure and want a framework you can actually read and modify
- Run IT operations and want an agent that builds institutional knowledge about your systems over time
The first version requires comfort with the command line. Enterprise forks are likely inevitable. But as of May 2026, for developers and technical operators, Hermes Agent is the most complete open-source answer to the memory problem that has limited AI agents since their emergence.
Hermes Agent is available at hermes-agent.nousresearch.com and on GitHub under NousResearch. MIT licence. All data stays on your machine.
Leave a Reply