How one developer keeps a self-learning AI agent in sync across two computers
A dev.to write-up shows how to share an agent's learned skills between machines without letting a tool silently overwrite your work.
The short version
- The agent saves each workflow it is taught as a 'skill'; the author had 172 across twenty categories.
- The installer refuses to overwrite the default profile, which protected the second PC's local skills.
- Live skills folder, repo copy and remote are three separate places that drift apart.
- Only about 8 MB of a 57 MB skills folder was real content; the rest was runtime clutter.
- Sync is triggered by hand and stops to ask on any conflict; no forced overwrites.
Why it matters
If you use an AI agent that learns from you, its knowledge is likely stuck on one computer. This post shows that the copying is easy and the risk is in deciding which copy wins. The same care applies to any tool that edits its own notes.
The agent that only knows one PC
As dev.to reported, a developer running an agent called Hermes ran into a problem familiar to anyone who works on two computers. The agent learns by example: each time he walks it through a multi-step job, such as publishing to a code server or troubleshooting a flaky Docker container, it saves the steps as a skill, a reusable procedure it loads the next time that task appears.
Over a few months that added up to 172 skills across twenty categories. He calls them his second brain. But the brain sat on one machine, so at his other PC the agent knew none of it.
The author says this affects any agent that learns locally, naming Claude Code, Codex and Cursor alongside Hermes. You either re-teach it everywhere, or let two copies drift until they disagree.
Three ways to move an agent
Hermes offers three official routes. One is a skills 'tap', a git repository (a version-tracked folder) of skill files. Another is a profile distribution, which packages the whole agent, including its personality file, settings, skills and scheduled jobs, as a git repo that another machine can update from with one command. The third is an export file for one-off moves or backups.
For keeping the same agent on two machines, he chose the distribution. He already ran his own git server, so it needed no new accounts or tools.
The guardrail that saved his work
The installer will not install into the default profile; you must name a new one. The author first saw this as an annoyance, then realized it was protection. His second PC already had an agent with project-specific skills, and installing over it, or forcing it with an overwrite flag, would have deleted that work.
So he installed alongside the existing agent and merged the two. His summary of the lesson is short: the merge is the product. Syncing, in his view, means reconciling two agents that each believe they are the original.
Three copies, not two
The subtler trap was where the files live. The distribution repo holds a snapshot of the skills folder; it is not the folder the agent actually writes to. So a sync involves three parties: the live folder, the repo copy, and the remote server.
A skill created after the last push exists only in the live folder. Had he treated the repo copy as the truth and synced from it, that skill would have vanished. So before any sync he compares the two folders with a standard diff command, and treats differences as normal rather than as errors.
Manual on purpose
He wants two-way sync, where both machines can edit skills, and wants to keep both versions and choose himself when the same skill changes in two places. A scheduled job cannot do that safely: it would either overwrite something or run while he is not watching.
- He says 'sync skills' and the agent runs the routine.
- It pulls from the remote in fast-forward-only mode, which refuses anything that would rewrite history.
- It compares folders, stages only real skill content, then pushes.
- It checks that the remote's latest commit matches the local one.
If anything is refused or a skill differs on both sides, it stops and asks. No forced pushes, no resets.
The clutter in the skills folder
When he first packaged the skills, the folder weighed 57 MB, but only about 8 MB was skill content. The rest was bookkeeping the agent's housekeeping process writes constantly, including a 40 MB index cache, backups and usage tracking. It changes daily and belongs to one machine, so he excluded it with a .gitignore file, which tells git what to skip.
The first real sync pushed a skill that existed only on one machine, and the same commit, edbf7ac, showed up on both ends. Existing skills were untouched.
One loose end: the agent's episodic memory (a record of past interactions, handled by a separate tool called mem0) does not travel through git. He says he should have decided early whether the machines share one memory store or keep their own.
What to watch
- Whether agent tools add built-in conflict handling so users don't have to build a manual routine.
- How agent memory stores, which sit outside git, will be shared or kept separate across machines.
Try it
If you use an agent that saves its own notes or skills, find where it stores them and check whether that folder is backed up or version-tracked. As general advice (not from the post), keep passwords and API keys out of any synced folder.
Sources and links
- Reportingdev.to: the article we sourced
First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.