NousResearch Launches Hermes Agent, an Open-Source Autonomous Agent With Persistent Memory and Self-Improving Skills

NousResearch, the open-source AI research group behind the Hermes model family, has released Hermes Agent, an MIT-licensed autonomous agent framework built around persistent memory, self-improving skills, and multi-platform deployment. The project went public on GitHub in early March and reached its v0.2.0 release on March 12.

The framework enters a crowded field that includes OpenClaw, LangChain, CrewAI, and NVIDIA’s NemoClaw. What distinguishes Hermes Agent is its focus on a closed learning loop: the agent creates skills from experience, refines them during use, and builds a persistent user model across sessions using Honcho for dialectic user modeling, according to the project’s GitHub README.

What It Does

Hermes Agent runs as a CLI application or as a messaging gateway that connects to Telegram, Discord, Slack, WhatsApp, Signal, and email from a single process. The official documentation describes 40+ built-in tools covering web search, browser automation, vision, image generation, text-to-speech, code execution, and subagent delegation.

The agent supports six terminal backends: local, Docker, SSH, Daytona, Singularity, and Modal. The serverless options (Daytona and Modal) allow the agent’s environment to hibernate when idle and wake on demand, which NousResearch positions as a cost advantage for persistent agents that don’t need constant compute.

Model selection is provider-agnostic. Users can connect to Nous Portal, OpenRouter (200+ models), OpenAI, or any OpenAI-compatible endpoint, and switch between them with a single command.

The Learning Loop

The core differentiator, per NousResearch’s product page, is what the team calls a “closed learning loop.” After completing complex tasks, the agent autonomously creates new skills and stores them for future use. It also maintains agent-curated memory with periodic nudges to persist important context, and supports FTS5 session search with LLM summarization for cross-session recall.

Skills follow the open AgentSkills.io standard, and can be installed from community repositories including ClawHub and LobeHub.

Where It Fits

The open-source agent framework space has fragmented rapidly in 2026. OpenClaw dominates consumer adoption. NVIDIA’s NemoClaw targets enterprise deployments with sandboxed execution and compliance controls. LangChain and CrewAI focus on developer tooling and orchestration.

Hermes Agent’s bet is on personalization and persistence: an agent that gets better the longer you use it, rather than starting from scratch each session. NousResearch’s existing reputation in the open-source community — the Hermes model series is widely used for fine-tuning and local deployment — gives the project a built-in audience among developers who prefer self-hosted, model-agnostic solutions.

The project includes a migration path for OpenClaw users via hermes claw migrate, per the GitHub repository, which suggests NousResearch views OpenClaw’s user base as a direct target.

Whether persistent memory and self-improving skills translate into meaningful retention advantages over competitors will depend on execution. The v0.2.0 release is early, and the project’s research-oriented roots (batch trajectory generation, Atropos RL integration) hint at ambitions beyond consumer tooling.