OpenClaw or Hermes Agent: which to run, and can you switch later?

In the ten days to October 21, OpenClaw shipped five stable releases and Hermes Agent six, so we also tested what a feature list can't show: moving a test agent from one to the other, in both directions.

You can leave in either direction. On one small test VM (not a DreamHost server) we moved an invented test agent from OpenClaw to Hermes with hermes claw migrate, then back with openclaw migrate hermes. The persona, memory files and skills came across both times. Both moves also stalled on something neither maker's guide mentions: on Docker installs each agent keeps its files private to its own user, so the other can't read them until you hand it a copy it owns. Hermes didn't report an error. It said Nothing to migrate from OpenClaw.

What each one used on our test VM

Our Linux test VM was arm64, with 2 vCPU and 4 GB of memory, the same size as our Stack 4 plan. We started each agent once from its maker's Docker files, with no AI model connected, and read its memory 2 and 10 minutes later. The full method is at the end of the page.

OpenClaw's gateway is the heavier of the two. It sat at 1,038 MiB after 2 minutes and 1,024 MiB after 10, and its memory high-water mark hit 1,737 MiB while it was starting up. Hermes, which runs its gateway and web dashboard in one container, read 583 MiB and then 557 MiB. Neither kept the processor busy. At 10 minutes, OpenClaw's median was 1.4% of one vCPU and Hermes's 0.9%.

OpenClaw's hardware FAQ lists 1 vCPU and 1 GB of memory as the minimum and recommends 2 GB or more. Size your server by the recommendation, since the idle gateway alone sat at about 1 GB before it had a channel, a chat or a key. Nous's Docker docs give Hermes a 1 GB minimum and recommend 2 to 4 GB, and say browser tools need at least 2 GB. On disk, the two unpacked images and their data came to about 10 GB, so you'll keep most of a 75 GB plan free.

Which one fits how you work

Hermes bills itself as self-improving. It turns a task it's finished into a skill it can reuse, and it searches its own past conversations. Its script-only cron jobs let you run a script on a schedule with no model call. Our test job fired on time with no key configured at all, so a job like an uptime check or a nightly report makes no model call. OpenClaw's scheduler came with three built-in jobs, and its payload docs list a --command job that runs a shell command with no model call. We didn't try one.

On OpenClaw's side, its own comparison page lists channels Hermes doesn't, among them Nextcloud Talk, Twitch, Urbit and three ways into Zalo. Community skills and plugins come from ClawHub, its public registry. The same comparison page describes its maker, the OpenClaw Foundation, as an independent nonprofit with no paid tier.

If you already run one of them and it does what you need, keep it. The import commands work, but you'd still re-pair WhatsApp, rebuild scheduled jobs and tidy skills, so switch only for a feature the other agent has. If you already pay for Claude Code, Hermes Agent vs Claude Code covers what an always-on agent adds. To run either one under a manager with budgets and tickets, see Paperclip vs OpenClaw or Hermes.

How to switch from OpenClaw to Hermes

Our test agent was an invented bakery assistant on OpenClaw 2026.10.4, with a persona, five memory entries, two days of notes and two skills. We moved it to a fresh Hermes v0.22.2 by following Hermes's migration guide, between the makers' Docker installs. The whole import run, container start included, took about 6 seconds once Hermes could read the files.

1. Save what the import won't bring

   OpenClaw keeps its state in ~/.openclaw and its workspace, where the memory files live, in ~/.openclaw/workspace. Its Docker file mounts that same host folder, plus login secrets from a separate ~/.openclaw-auth-profile-secrets folder, so copy that too. Copy both off the server, because a Git copy of your workspace alone misses session history and logins.

   Save your openclaw cron list output too. Hermes's docs say scheduled jobs are archived for you to rebuild, but our import didn't pick up OpenClaw's three built-in jobs at all (No cron configuration found). If Hermes will run on a second server, copy the folder there with rsync -a and run the remaining steps on that server.

2. Give Hermes a copy it can read

   OpenClaw's files are readable only by its own user, uid 1000. Hermes runs as uid 10000 in its image. If your setup's users differ, you can check each folder's owner with ls -ln.

   bashCopy

   sudo cp -a ~/.openclaw /srv/openclaw-copy
   sudo chown -R 10000:10000 /srv/openclaw-copy

3. Stop Hermes and preview the import

   We stopped the Hermes container first. Then we ran the import from its image, with the copy mounted read-only and Hermes's data folder at /opt/data. The preview lists what it would import and what it would skip. A dry run doesn't change anything. Add --workspace-target to tell it where your AGENTS.md goes, or it stays behind.

   bashCopy

   docker stop hermes
   docker run --rm \
     -v /srv/openclaw-copy:/src/openclaw:ro \
     -v ~/.hermes:/opt/data \
     nousresearch/hermes-agent:v0.22.2 \
     claw migrate --dry-run \
     --source /src/openclaw \
     --workspace-target /opt/data/workspace

   On a native install, run the same preview as hermes claw migrate --dry-run with --source /srv/openclaw-copy and --workspace-target ~/.hermes/workspace.

4. Apply it with --overwrite

   A fresh Hermes counts its own default persona and model as two conflicts. It wouldn't apply until we added --overwrite. It writes a restore-point zip first. You can restore it with hermes import <zip>. API keys come across only with --migrate-secrets, and only the ones on Hermes's allowlist.

   bashCopy

   docker run --rm \
     -v /srv/openclaw-copy:/src/openclaw:ro \
     -v ~/.hermes:/opt/data \
     nousresearch/hermes-agent:v0.22.2 \
     claw migrate --preset full --overwrite --yes \
     --source /src/openclaw \
     --workspace-target /opt/data/workspace
   docker start hermes

5. Check memory and finish the parts that don't move

   Hermes's docs say imported memory is deduped, but an entry we'd repeated in a daily note showed up twice, so read your MEMORY.md. Start a new session, because imported skills and memory won't load in the session you're in. Re-pair WhatsApp with hermes whatsapp. Recreate scheduled jobs with hermes cron create. Then stop OpenClaw's gateway so only Hermes answers your bots, and once you're happy, hermes claw cleanup renames the leftover OpenClaw folders to .pre-migration/.

Our OpenClaw-to-Hermes import moved 8 items and skipped 25, and the restore zip was 1.2 MB.

How to switch from Hermes to OpenClaw

OpenClaw's import needs a fresh OpenClaw setup, so run it before you build up state on the new side, or apply with --overwrite once you've read the plan. Per OpenClaw's guide, it copies the persona and AGENTS.md and brings skills with their support files. Hermes's MEMORY.md and USER.md go on the end of OpenClaw's own memory files. Sessions, scheduled jobs, logs, state databases and channel pairings go into a report folder for review and are never loaded.

1. Back up OpenClaw yourself and stop its gateway

   OpenClaw writes its own backup before an import, but on its Docker install our backup landed inside the one-off CLI container and was gone when the container exited.

   bashCopy

   sudo cp -a ~/.openclaw ~/openclaw-before-import
   docker compose stop openclaw-gateway

2. Give OpenClaw a copy it can read

   Hermes's files are private to uid 10000, so OpenClaw hit the same wall. Its own user is uid 1000.

   bashCopy

   sudo cp -a ~/.hermes /srv/hermes-copy
   sudo chown -R 1000:1000 /srv/hermes-copy

3. Preview and apply from the CLI container

   --from points the import at the copy, mounted into the one-off container from OpenClaw's Docker file.

   bashCopy

   docker compose run --rm -v /srv/hermes-copy:/src/hermes \
     openclaw-cli migrate hermes --dry-run --from /src/hermes
   docker compose run --rm -v /srv/hermes-copy:/src/hermes \
     openclaw-cli migrate apply hermes --yes --from /src/hermes

   If you run the import interactively, it asks whether to bring your keys and logins, with yes already selected. The --yes form imports non-secret state only unless you add --include-secrets.

4. Start the gateway and run doctor inside it

   The documented docker compose run --rm openclaw-cli form refused status, doctor and skills list while the gateway was running, so use exec for your day-to-day commands.

   bashCopy

   docker compose start openclaw-gateway
   docker compose exec openclaw-gateway \
     node dist/index.js doctor

5. Tidy skills and bring AGENTS.md over by hand

   OpenClaw copied 60 skills, our 2 plus the 58 skill folders Hermes had synced from its bundle, and doctor warned that they shadow 11 of OpenClaw's own, including github and apple-notes. Delete the ones you don't want from ~/.openclaw/workspace/skills. OpenClaw's --skill flag imports only the skills you name, which we didn't test. AGENTS.md didn't come back at all. Our copy sat in Hermes's workspace folder, where the first import put it, and OpenClaw only archives that folder.

   bashCopy

   sudo cp /srv/hermes-copy/workspace/AGENTS.md \
     ~/.openclaw/workspace/
   sudo chown 1000:1000 ~/.openclaw/workspace/AGENTS.md

The preview listed 81 items and no conflicts, and the import run took about 4 seconds. The persona, the model setting and both memory files arrived, with Hermes's entries under an Imported from Hermes marker.

Side by side

We read both projects' docs on October 16, 2026, when the latest releases were OpenClaw 2026.10.4 and Hermes Agent v0.22.2, and re-checked them on October 21, when the latest were 2026.10.5 and v0.22.4. The memory, disk, time and no-key rows come from one run of each maker's Docker files on our arm64 test VM.

The channel rows come from OpenClaw's comparison page, which also warns that the two catalogs' counts are not directly comparable. Both cover Telegram, WhatsApp, Slack, Discord and Signal. Hermes reaches iMessage through BlueBubbles. Check that your own app is in the Hermes list or OpenClaw's before you choose on channels.

Running both while you decide

With the two started together on our 4 GB VM, OpenClaw read 1,140 MiB after 10 minutes and Hermes 557 MiB, and the VM still had 2,081 MB available. That's about 115 MiB more than OpenClaw read alone, so plan on 1.0 to 1.15 GB for it at idle. We didn't connect a model, and we'd expect a working agent to use more than these idle figures. Their default ports don't clash: OpenClaw's compose file uses 18789, 18790 and 3978, and Hermes's uses 8642 and 9119.

Give each of your agents its own bot. Hermes's FAQ says each messaging platform needs exclusive access to a bot token, so a second gateway polling the same Telegram bot won't connect. If you imported an OpenAI or Copilot login from one agent to the other, sign in again on one side before you run both. OpenClaw's migration notes say the two shouldn't share one refresh grant, the saved sign-in each uses to stay logged in.

Hermes's dashboard login

Hermes's username and password login comes from HERMES_DASHBOARD_BASIC_AUTH_USERNAME and HERMES_DASHBOARD_BASIC_AUTH_PASSWORD. Set both before you start its Docker example.

How we measured

DreamHost ran the test on October 16, 2026. The VM ran Ubuntu 24.04.4 through Colima on an Apple-silicon Mac, with 2 vCPU, 3,904 MB of usable memory, no swap and Docker 29.5.2. OpenClaw 2026.10.4 started from the setup script and compose file published with that release, with onboarding skipped by the script's own switch. Hermes Agent v0.22.2 started from the compose example on its Docker page, with the username and password login added so the dashboard would run. We'd downloaded the images first. We ran each setup once, so each memory figure is the median of 12 docker stats readings taken 5 seconds apart within that one run, rounded to the nearest MiB. OpenClaw's second start, in the both-at-once run, read about 115 MiB higher, which gives a sense of how far a repeat could move. The start-up peak is the high-water mark the kernel keeps for each container, a different counter from the idle readings. The switch timings cover the whole one-off container run, start-up included.

We re-checked every doc, version and price on this page on October 21, 2026, and didn't rerun the test on the newer releases. On October 22 we checked every DreamHost plan, price and product statement here against the live OpenClaw and Hermes Agent VPS Hosting pages. We didn't give either agent a model key, bot token or account. Nobody chatted with them, so the figures show each agent sitting ready. The switch used invented test data and no secrets. We didn't test answer quality, model cost, a week of real use, live chat channels, moving keys or session history, either maker's native installer, or the switch on our own images. Our OpenClaw and Hermes images are set up differently from the makers' Docker files, so your numbers will differ.

OpenClaw and Hermes Agent FAQs