Moving from OpenClaw to Hermes sounds like a tooling decision. It is not. It is a state migration with a new interface attached.
Keepmyclaw currently supports OpenClaw and Hermes backup and restore only. That scope matters. This is not a generic agent migration playbook, not Claude Code state loss advice, and not a promise that every local coding assistant can be restored the same way. The target is OpenClaw and Hermes Agent workspaces where useful operating context lives in files, memory, skills, cron jobs, scripts, configs, and credentials after client-side encryption.
The operator thesis is simple: migrate only after you can roll back. If Hermes starts clean but your OpenClaw workspace cannot be restored, you did not migrate. You performed a ritual sacrifice with a nicer dashboard.
Why this migration is easy to underestimate
OpenClaw and Hermes overlap enough to make the move feel safe. Both are local operator tools. Both can involve long lived workspaces, memory, scripts, credentials, and scheduled work. Both attract users who do not want their agent setup trapped in a web chat box.
That overlap is exactly where people get sloppy.
The risky state is rarely the obvious repo. It is the stuff around the repo. A local persona file. A one off recovery script. A cron job prompt that took ten iterations to stop sounding like a brochure. A credentials note that says which token is safe to rotate and which one will break three jobs. A small state ledger that prevents a marketing agent from reposting the same thing twice.
Lose those and the new Hermes setup may still boot. It may even look productive. Then a scheduled job fires with the wrong toolset, a restore script points at the old path, or the agent confidently repeats work that the old ledger would have blocked. Small misses become expensive when nobody remembers why the old setup worked.
The backup contract before you move
Set a migration contract before touching the old workspace. Numbers are useful because they leave less room for vibes.
Use these defaults unless your setup has stricter needs:
- Recovery point objective: 15 minutes for active OpenClaw or Hermes workspaces
- Recovery time objective: 60 minutes to restore the old OpenClaw workspace onto a clean machine
- Migration rollback window: 7 days after Hermes becomes the primary workspace
- Snapshot cadence during migration week: every 15 minutes while active
- Pre-migration retention: 14 daily snapshots and 8 weekly snapshots
- Post-migration overlap: keep both OpenClaw and Hermes snapshots for 30 days
- Restore drill cadence: one full drill before migration, one drill within 72 hours after migration
- Cron verification target: 100 percent of migrated jobs listed with schedule, prompt, skills, script, and delivery target intact
- Skill verification target: expected skill count, file sizes, and checksums match the manifest
- Credential rebind window: 60 minutes, with secret values never printed into logs
- Maximum unverified snapshot age: 30 days
These are not laws. They are tripwires. If you cannot meet them, slow down. The migration is already telling you where it will hurt.
What to back up from OpenClaw first
Start with the workspace files. Repos, docs, notes, generated assets, local config, and any state files that the agent reads or writes. Include hidden files if they change behavior. Hidden files love becoming important after you forget them.
Back up memory next. That means durable memory files, persona notes, skills, prompt rules, and any local playbooks that shape how the agent acts. Do not trust names alone. Record paths, sizes, checksums, and last modified times. If two files are called memory.md, the manifest should tell you which one mattered.
Then back up scheduled work. For Hermes, cron jobs are first class operational state. For OpenClaw-style setups, the same rule applies even if scheduling is implemented differently. Preserve job names, prompts, schedules, enabled tools, attached skills, scripts, context dependencies, delivery targets, repeat rules, and dedupe ledgers.
Back up scripts and their assumptions. A script without its environment is archaeology. Capture runtime versions, package manifests, expected working directories, safe test commands, and which scripts are read only versus public or destructive.
Back up credentials only after client-side encryption. Store enough metadata to rebind access safely: provider name, account label, scope, expected file path, token presence, and token length. Never store raw secrets in reports, PR descriptions, or migration notes. The whole point is disaster recovery, not turning GitHub into a credential piñata.
The clean rollback test
Before moving primary work to Hermes, restore OpenClaw onto a clean target. Not the same laptop. Not the same home directory. A clean target.
The test is boring by design:
- Install the required runtime and OpenClaw or compatible local setup.
- Restore the selected snapshot using the manifest.
- Verify workspace paths and file checksums.
- Load memory, skills, prompts, and local playbooks.
- Rebind encrypted credentials with safe read only checks.
- List scheduled jobs or their OpenClaw equivalent.
- Run one harmless read only job or script.
- Confirm the restored workspace can perform one real operator task without missing context.
If this takes more than the RTO, fix the backup before migration. Do not explain it away. Future you will not become calmer during an outage.
Failure scenarios worth catching
Hermes works but the old workspace cannot roll back
The new setup may feel fine for a day. Then you discover one project still depends on an OpenClaw path, plugin, or memory file. If the old workspace is not restorable, rollback becomes manual reconstruction.
Mitigation: keep a verified OpenClaw snapshot and a clean-machine restore note for at least 30 days after the move. Treat rollback as part of the migration, not an embarrassing side quest.
Cron jobs migrate without their guardrails
A job title and schedule are not enough. The dangerous parts are often inside the prompt: public action caps, silence rules, allowed accounts, forbidden topics, dedupe ledgers, and delivery targets.
Mitigation: export the full job definition and compare it after restore. Run safe jobs only. Do not live test public posting, email sending, checkout changes, or deployment jobs unless that boundary is explicitly approved.
Skills restore but persona files do not
A skill can load correctly while the local persona file it references is gone. The agent still answers. It just sounds wrong, picks stale topics, or violates a hard voice rule.
Mitigation: treat referenced files as dependencies. The manifest should include skills, persona files, playbooks, and any state examples the skill reads.
Credentials exist but scopes changed
A restored token file can be present and still useless. The account may be wrong, the scope may be too narrow, or the token may have expired.
Mitigation: verify credentials through safe metadata endpoints. Check account label, scope, and length only. Do not print secrets. Do not mutate auth while testing a backup unless the migration plan explicitly allows it.
Ledgers reset and the agent repeats work
Dedupe ledgers look disposable until they are gone. Then a social job reposts a source, a blog job drafts a duplicate angle, or a recovery job contacts the same lead twice.
Mitigation: include ledgers under the backup scope. Validate counts before and after restore. For public-action ledgers, append only after the public action is verified, not after a draft or partial run.
What changes after Hermes becomes primary
Once Hermes is primary, keep backing up both systems during the overlap window. The old OpenClaw workspace becomes rollback state. The new Hermes workspace becomes production state. Both matter until you prove the new setup carries the full workload.
Run a post-migration restore drill within 72 hours. Restore Hermes onto a clean target. Confirm skills, memory, cron jobs, scripts, configs, encrypted credentials, and dedupe ledgers came across. Then run one safe read only job and check the output path.
Only retire the OpenClaw rollback snapshot after the overlap window closes and the Hermes restore drill passes. If the drill fails, the answer is not optimism. The answer is another snapshot and a fixed manifest.
The migration checklist
Before migration:
- create a full OpenClaw snapshot
- generate a manifest with paths, sizes, checksums, versions, and timestamps
- include memory, skills, persona files, scripts, configs, cron jobs, and ledgers
- encrypt credentials client side before backup
- restore OpenClaw onto a clean target and record the result
During migration:
- keep snapshots running every 15 minutes while active
- copy only known state, not mystery folders
- map old paths to new Hermes paths in writing
- verify every migrated cron job and skill dependency
- avoid public or destructive live tests
After migration:
- run a Hermes clean-machine restore within 72 hours
- keep OpenClaw rollback snapshots for 30 days
- compare ledger counts and job counts before retiring the old setup
- update the restore note with the actual commands that worked
The move from OpenClaw to Hermes should feel boring. Boring means the old workspace can come back, the new workspace can be restored, and nobody has to remember which random file made the agent useful.
That is the bar. Anything below it is just changing tools and hoping the floor holds.