Hermes Agent Migration Case Study

Source: Sathish Raju, I Switched from OpenClaw to Hermes Agent. Here’s What Nobody Told Me (Medium, 2026-04-22): https://medium.com/@sathishkraju/i-switched-from-openclaw-to-hermes-agent-heres-what-nobody-told-me-5f33a746b6ca

This is a summary in my own words; read the original for the full text, diagrams and code samples.

Overview

OpenClaw (Peter Steinberger, late 2025) and Hermes Agent (Nous Research, released 2026-02-25) are both open-source, self-hosted agent frameworks that work with any major model provider and connect to common messaging apps. The article compares them on one main question: does the agent keep anything from a finished task? OpenClaw starts each task from the same baseline. Hermes adds a step after each task that turns the experience into a reusable skill, so repeated work should get cheaper.

Why the comparison came up

Figures as reported in the article, early April 2026:

OpenClaw’s growth — about 345,000 GitHub stars, 13,000+ community skills on the ClawHub marketplace, 24 messaging integrations.
Its March 2026 security problems:
- Nine CVEs in four days, the worst rated CVSS 9.9; one example, CVE-2026-25253, let anyone on the network pull every stored API key through an unauthenticated endpoint.
- 341 of the first 2,857 ClawHub skills audited were malicious (about 12%).
- 135,000+ instances were reachable from the public internet.
Diagnosis — a tool built for one person on one machine became a networked service, and its local assumptions did not survive.

How Hermes learns

After a complex task, Hermes runs a reflection pass with three parts:

Skill creation — writes a markdown skill file of the steps that worked, in the portable agentskills.io format, on top of 118 bundled skills.
User modeling — keeps a cross-session profile of format and detail preferences, past decisions and feedback, so it asks fewer clarifying questions.
Self-evaluation — tests whether a new skill generalizes beyond the case that produced it, and refines it before adding it to the library.

On a similar later task it consults its own library first. The article claims roughly 40% faster completion once 20+ skills exist in a domain; no method is given. The core contrast: OpenClaw’s skills are written by people and installed; Hermes writes its own.

Memory and setup

Pluggable memory (v0.10.0) — SQLite with a full-text episode archive by default; six third-party backends such as Honcho or vector stores.
Less transparent — OpenClaw keeps each memory as an editable file; Hermes’s archive is hard to read, and hermes memory inspect only partly helps.
Install — pip install hermes-agent or from source, then set the model provider and memory options in the config under ~/.hermes/.
Migrate — hermes claw migrate --source ~/.openclaw imports persona, memory, skills, config and API keys; preview with --dry-run, then check with hermes skills list and hermes memory status.

Note: The learning loop is off by default. Without enabling memory and skill generation, Hermes is an ordinary single-session agent; the author blames many unimpressed reviews on this.

Security and limitations

Security — no agent-specific Hermes CVEs as of April 2026, plus container hardening, subagent namespace isolation and no marketplace to poison. The author’s caveat: a two-month-old project has had little exposure, so this is encouraging, not proof; review either framework before enterprise use.
Not a coding tool — Claude Code, Cursor or Windsurf are better for code; Hermes suits research, analysis and recurring workflows.
Fewer integrations — six messaging platforms versus 24+, by design.
Young — about 10 releases versus OpenClaw’s 82.

Which to choose

Situation	Choice
The same kinds of task daily, with months to let skills accumulate	Hermes
Many messaging channels, team-wide deployment, file-based memory, or existing ClawHub investment	OpenClaw
Complex setups	Both: OpenClaw routes across channels; Hermes executes the workflows where learning pays off

The “use both” pattern is the multi-agent lesson here: one layer coordinates, another specializes and improves. Costs are similar (the article estimates $30–65/month in API calls for Hermes on a $5 VPS, $40–80 for OpenClaw on self-managed Docker), so the decision is about trade-offs, not price.