Hermes Agent is a self-improving artificial intelligence agent developed by Nous Research, distinguished by its built-in learning loop that enables continuous skill acquisition, refinement, and knowledge persistence. It is designed for high accessibility and flexibility, supporting deployment on diverse infrastructure from $5 Virtual Private Servers (VPS) to GPU clusters or serverless environments like Daytona and Modal, which offer cost-effective hibernation when idle. The agent operates independently of a local machine, allowing interaction via various messaging platforms such as Telegram, Discord, Slack, WhatsApp, and Signal, in addition to a rich terminal user interface (TUI).

A core technical strength of Hermes Agent lies in its model agnosticism. It supports a wide array of Large Language Models (LLMs) and providers, including Nous Portal, OpenRouter (offering access to over 200 models), z.ai/GLM, Kimi/Moonshot, MiniMax, OpenAI, and custom API endpoints. Users can seamlessly switch between models and providers using CLI commands like hermes model, ensuring no vendor lock-in.

The agent's "closed learning loop" is central to its methodology:

• Agent-Curated Memory: Hermes maintains a persistent memory system that is actively curated by the agent. This includes user profiles, conversation histories, and critical information, periodically nudged for consolidation.
• FTS5 Session Search: For cross-session recall, the agent employs FTS5 (Full-Text Search Engine for SQLite) to search past conversations, leveraging LLM summarization to extract and present relevant insights.
• Honcho Dialectic User Modeling: This refers to an approach for building a deepening model of the user over time, facilitating personalized interactions and contextual understanding across sessions.
• Autonomous Skill Creation: When faced with complex tasks, the agent can autonomously create new skills from experience. These skills encapsulate learned procedures or problem-solving methodologies.
• Skill Self-Improvement: Beyond creation, existing skills are designed to improve themselves during use, suggesting an iterative refinement process based on performance feedback and subsequent applications.
• Knowledge Persistence: The system actively nudges itself to persist acquired knowledge, translating transient experiences into long-term memory or actionable skills.

Hermes Agent incorporates robust tooling and architectural features:

• Toolset System: It provides access to over 40 tools, organized via a flexible toolset system, and is compatible with the agentskills.io open standard.
• Terminal Backends: It supports six different terminal backends (local, Docker, SSH, Daytona, Singularity, Modal) for executing commands and tools in various environments, with serverless persistence options from Daytona and Modal.
• Scheduled Automations: A built-in cron scheduler allows for unattended execution of tasks, such as daily reports, nightly backups, or weekly audits, delivered to any supported platform in natural language.
• Subagent Parallelization: The agent can spawn isolated subagents for parallelizing workstreams, enabling complex tasks to be broken down and executed concurrently.
• Remote Procedure Call (RPC) for Tools: Python scripts can call tools via RPC, collapsing multi-step pipelines into single, "zero-context-cost" turns, enhancing efficiency and reducing computational overhead.
• Research Capabilities: The project supports research into AI agents, offering functionalities like batch trajectory generation, integration with Atropos Reinforcement Learning (RL) environments, and trajectory compression, which are crucial for training next-generation tool-calling models.

The user interface includes a full TUI with advanced features like multiline editing, slash-command autocompletion, conversation history, and real-time streaming of tool outputs. Command-line utilities include hermes for interactive CLI, hermes model for LLM configuration, hermes tools for managing tools, hermes config set for general settings, hermes gateway for messaging platform integration, hermes setup for initial configuration, and hermes update for version management. The system also supports migration from OpenClaw, importing personas (SOUL.md), memories (MEMORY.md, USER.md), user-created skills, command allowlists, messaging settings, and API keys.

The project promotes community contributions, adhering to a structured development setup, code style, and PR process, using uv for dependency management and pytest for testing. It is released under the MIT License.