Mythos

The Hermes Glossary is a collection of terms used in 📝Hermes Agent, Nous Research's open-source self-improving autonomous agent — spanning its runtime architecture, learning loop, skills system, multi-agent orchestration, and messaging gateway.

Glossary

A

  • agentskills.io — open standard for portable, shareable Markdown skill documents that Hermes implements and contributes to.
  • AnthropicTransport — built-in provider transport for routing requests through the Anthropic API while sharing Hermes' streaming, retry, and cache logic.
  • Atropos — Nous Research's reinforcement learning environment for training tool-calling models; Hermes integrates with Atropos through the Tinker-Atropos submodule.
  • Auxiliary models — per-task model overrides for compression, vision, session search, and title generation, configured under hermes model → "Configure auxiliary models."
  • Autonomous Curator — built-in hermes curator process that reviews agent-created skills, consolidates overlap, archives stale entries, and protects pinned skills.

B

  • Batch trajectory generation — built-in tooling that captures agent runs as training data for fine-tuning tool-calling models with Atropos.
  • BedrockTransport — provider transport for routing requests through AWS Bedrock via the Converse API.

C

  • Camoufox — local anti-detection browser backend for stealth browsing, with persistent sessions and VNC URL discovery for visual debugging.
  • ChatCompletionsTransport — provider transport for any OpenAI-compatible Chat Completions endpoint.
  • Checkpoints v2 — v0.13.0 state-persistence rewrite with real pruning, replacing the earlier checkpoint manager.
  • claw migrate — hermes claw migrate subcommand that imports settings, memories, skills, and API keys from an existing ~/.openclaw install.
  • Closed learning loop — Hermes' core architectural claim: solve → document → retrieve → improve → repeat, executed across sessions without human prompting.
  • /compress — slash command that triggers guided context compression with an optional focus topic to preserve.
  • Container hardening — sandbox defaults including read-only root filesystems, dropped capabilities, namespace isolation, and a pre-execution scanner for terminal commands.
  • Credential pool — same-provider multi-API-key configuration with thread-safe least_used rotation and automatic failover on 401 responses.

D

  • Dashboard — local browser-based web interface for managing Hermes (settings, sessions, skills, gateway), with a plugin system for custom tabs and live theme switching.

F

  • Fast mode — /fast toggle routing requests through OpenAI Priority Processing and Anthropic's fast tier for lower latency on supported models.
  • FTS5 session search — SQLite full-text search across every past session, paired with LLM summarization for cross-session recall.

G

  • Gateway — single process (hermes gateway) that runs every messaging integration in parallel, keeping conversations continuous across 20+ platforms.
  • /goal — v0.13.0 command that locks the agent on a target across turns, implementing the Ralph loop pattern.

H

  • Hermes-3 — the LLaMA-3.1 fine-tune from Nous Research, tuned for tool-calling accuracy and long-range planning; one of many models the agent can run on, not a dependency.
  • HermesClaw — community WeChat bridge that runs Hermes Agent and OpenClaw on the same WeChat account.
  • Honcho — third-party dialectic user modeling library Hermes integrates as the reference pluggable memory provider, building a persistent evolving model of the user across sessions.

K

  • Kanban — v0.13.0 durable multi-agent board with heartbeat, reclaim, zombie detection, auto-block on incomplete exit, per-task retries, and hallucination recovery.
  • KANBAN_GUIDANCE — system prompt fragment auto-injected into every Kanban worker, generated from agent/prompt_builder.py.
  • Kanban worker — isolated subagent operating against the Kanban board with its own conversation, terminal, and Python RPC scripts, used for zero-context-cost pipelines.

L

  • 📝LLaMA — Meta's open-weight model family; Hermes-3 is built on LLaMA 3.1.
  • /llms.txt — Hermes docs endpoint providing a curated ~17 KB index of every documentation page, safe to load into an LLM context.

M

  • MCP — Model Context Protocol, the open standard for exposing tools, resources, and prompts to LLM agents; Hermes connects to any MCP server for extended tool capabilities.
  • Messaging Gateway — see Gateway.

N

  • Nomos — one of Nous Research's foundation models, sibling to Hermes and Psyche.
  • Nous Portal — Nous Research's first-class inference provider, offering 400+ models via inference-api.nousresearch.com/v1.
  • Nous Research — the model-training lab that builds Hermes Agent and the Hermes, Nomos, and Psyche model families.
  • Nous Tool Gateway — paid Nous Portal feature routing web search, image generation, TTS, and browser automation through Nous infrastructure.

O

  • 📝Ollama — local LLM runtime supported as a Hermes provider for zero-API-cost setups.
  • 📝OpenClaw vs. Hermes — comparison memo positioning Hermes against the larger but security-troubled OpenClaw ecosystem.
  • 📝OpenRouter — multi-model gateway providing 200+ models through a single API, supported as a Hermes provider.
  • Orchestrator — subagent role that can spawn its own workers, with configurable max_spawn_depth for nested delegation.

P

  • Pluggable context engine — hermes plugins slot for swapping in custom context management — filtering, summarization, or domain-specific injection — that controls what the agent sees each turn.
  • Pluggable memory provider — extensible plugin interface for memory backends; the built-in store is the default, with Honcho as the reference third-party provider.
  • Pre-execution scanner — Hermes' static-analysis pass that inspects terminal commands before they run inside a sandbox backend.
  • Persistent memory — agent-curated long-term storage across all sessions, surfaced via FTS5 search and LLM summarization.
  • ProviderTransport — abstract base class in agent/transports/ that every concrete transport (Anthropic, Chat Completions, Responses API, Bedrock) implements.
  • Psyche — Nous Research model in the same family as Hermes and Nomos.

R

  • Ralph loop — execution pattern implemented by /goal that keeps the agent locked on a target across turns until the goal is met.
  • ResponsesApiTransport — provider transport for the OpenAI Responses API.

S

  • Sandbox backend — execution environment for terminal commands and code; Hermes ships six (local, Docker, SSH, Daytona, Singularity, Modal).
  • Self-improving skill — skill document patched in place during use when the agent encounters information that contradicts or extends it.
  • Skill document — Markdown file with YAML frontmatter (name, description, version, platforms, metadata) following the agentskills.io standard; Hermes generates one autonomously after complex tasks (typically 5+ tool calls).
  • Skill Hub — community marketplace at agentskills.io for official optional and community-contributed skills, installed via hermes skills install.
  • Slash command — user-typed /foo that loads a skill or built-in command; works identically in the CLI and every gateway-connected platform.
  • /steer — mid-run nudge that injects a note the running agent sees after its next tool call, without interrupting the turn or breaking prompt cache.
  • Subagent — isolated agent spawn with its own conversation, terminal, and tools, dispatched for delegation or parallel work; can take the orchestrator role to spawn further workers.

T

  • Tenacity Release — codename for v0.13.0, themed around durable execution (Kanban, /goal, Checkpoints v2, gateway auto-resume).
  • Terminal UI — Hermes' interactive CLI with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output.
  • Tinker-Atropos — RL training submodule bundled with Hermes that wires agent trajectories into Atropos training environments.
  • Trajectory — full sequence of reasoning steps, tool calls, and outputs from one task; used as the unit of training data for Atropos.
  • tui_gateway — Python JSON-RPC process that backs the React/Ink interactive CLI launched with hermes --tui.

W

  • watch_patterns — background process monitoring that triggers real-time alerts when output matches configured patterns, eliminating the need to poll for completion.
  • Webhook direct-delivery mode — webhook subscription mode that forwards payloads straight to a platform chat without going through the agent, enabling zero-LLM push notifications.

Contexts

Created with 💜 by One Inc | Copyright 2026