This post compares two MIT-licensed open-source AI agent frameworks that both exceeded 10K GitHub stars but pursue entirely different design philosophies: OpenClaw is a multi-platform message gateway + tool orchestration engine, while Hermes Agent is a self-evolving long-term companion. The choice depends on whether you need to "scale channels horizontally" or "grow memory vertically." Content is based on the original post, official docs, and community materials.
Key points
- Positioning: OpenClaw's motto is "Give AI a computer and let it work" (Peter Steinberger, community-driven, early 2025, 333K+ stars). Hermes is "The agent that grows with you" (Nous Research, released Feb 2026, 66K+ stars in 2 months). OpenClaw's north-star metric is platform/tool coverage; Hermes's is how many repeated instructions it eliminates after 3 months of use.
- Runtime architecture:
- OpenClaw treats 4 runtimes as first-class citizens: PI (Process Intelligence), Codex CLI, Claude CLI, and ACP (Agent Control Protocol). A Node.js Gateway (default
127.0.0.1:18789) routes messages from all channels into sessions, runs the agent loop against external LLM APIs, and executes tool calls. Sessions run serially; multi-agent routing and zero-context-cost subagent spawning are supported. - Hermes uses a single native agent loop with a built-in learning cycle:
user interaction → behavior logging → result evaluation → strategy optimization → skill consolidation. Tasks with 5+ tool calls are automatically distilled into Skills (Markdown + YAML frontmatter) with progressive disclosure; a global performance self-review runs every 15 tasks. - Memory systems:
- OpenClaw is file-first and fully transparent: working memory (
SOUL.md,USER.md,AGENTS.md), daily archives (memory/YYYY-MM-DD.md), SQLite + vector embeddings for semantic search, human-curatedMEMORY.md, and optional LightRAG / GraphRAG / PostgreSQL. Memory is treated as the user's property (Git-friendly, auditable), but there is no automatic pruning. - Hermes uses SQLite + FTS5 with 8 pluggable providers (honcho, mem0, hindsight, etc.): cache-aware hot memory, LLM-summarized full-text session search, Honcho dialectic user modeling, auto-created skills in
~/.hermes/skills/, a meta-memory layer, and a background Curator that evaluates/merges/upgrade-prunes rather than simply deleting. It freezes system-prompt snapshots for cache-aware billing. Trade-off: memory is opaque to the user. - Authentication: OpenClaw centralizes all credentials in
auth-profiles.json(unified rotation/audit, but single-point breach risk), with localhost-only Gateway binding, tool sandboxing + Docker isolation, prompt-injection defenses, and human approval for sensitive actions. Hermes delegates auth to the Codex CLI when using OpenAI, with each other provider managing its own keys, plus command approval, container isolation, DM pairing, and a conservative default posture. - Channels: OpenClaw covers 27+ messaging platforms (WhatsApp, Telegram, Discord, Slack, Signal, MS Teams, Google Chat, Feishu, WeCom, QQ, Weibo, etc.) at message level. Hermes covers 19+ (Telegram, Discord, Slack, WhatsApp, Signal, Email, Home Assistant, etc.) plus deep IDE integration (VS Code / Zed / JetBrains) via ACP — editor-level code, terminal, and filesystem access.
- Plugin ecosystems: OpenClaw's ClawHub hosts 5,700+ community skills in Markdown + YAML format (Claude Code / Cursor compatible), mostly hand-written. Hermes ships 85 built-in skills and 520+ community skills compatible with the agentskills.io open standard (interoperable with OpenClaw), and its key advantage is automatic skill creation and self-improvement after complex tasks.
- Who should choose which:
- OpenClaw: multi-platform automation (WhatsApp + Telegram + Slack + Email), full transparency/auditability, manual memory curation, Node.js/TypeScript stacks — e.g., freelancers, personal life automation, small-team bots.
- Hermes: long-term companion agents, repetitive workflows, code assistance, research/reporting, Python/research stacks — e.g., long-running dev projects, daily data-collection reports, creator automation.
- They can coexist: the most practical setup is "OpenClaw as the message gateway, Hermes as the backend brain" — OpenClaw ingests channel messages, forwards complex tasks to Hermes (leveraging its learning loop), and redistributes results. Hermes even ships a
hermes claw migratecommand to import OpenClaw's SOUL.md, MEMORY.md, skills, and API keys. - Philosophy: OpenClaw treats an agent as an LLM wrapper with tools, memory as user property, skills as hand-written manuals, and time as present-task execution — an "operating system" model. Hermes treats an agent as a self-evolving digital entity with system-managed memory, auto-consolidated procedural memory, and long-term compounding — a "living organism" model.
- OpenClaw docs: https://docs.openclaw.ai
- Hermes Agent docs: https://hermes-agent.nousresearch.com/docs
- Nous Research GitHub: https://github.com/NousResearch/hermes-agent
- OpenClaw GitHub: https://github.com/openclaw/openclaw