I came across Hermes some time back. It markets itself as a personal AI assistant that remembers your preferences and handles recurring tasks on your behalf. The memory angle caught my attention, most AI tools are stateless by default, so the idea of something that accumulates context over time and acts on it autonomously seemed worth exploring.
Setup
Getting Hermes into a useful assistant with permanent memory requires solutions such as Obsidian. You point it at your vault, it reads and writes notes there, and that becomes its persistent memory layer. From there, I configured a few tasks to run on a cron schedule, a daily digest of sorts, pulling together some information I would otherwise check manually.
Initial impressions were positive. The setup process was smoother than I expected and the tasks ran without issues.
The Credit Consumption Problem
The first thing I noticed after a couple of days was the credit usage. It was higher than I anticipated for what amounted to a few lightweight scheduled tasks. This is perhaps an unavoidable trade-off, Hermes has to figure out the environment, install packages and so on. Compared to just vibecoding the script with Claude, it does more work and thus consumes more credits.
What I did not realise at first was the extent of this. Every time a task ran, the output was being passed back through Hermes for processing. So a task that runs four times a day is not just running a script four times, it is making four AI API calls, each consuming credits, regardless of whether the output required any intelligence to interpret. I noticed unexplained credit consumption on a day I had not interacted with Hermes at all. After some digging, this was the explanation.
The Abstraction Tax
The other thing that became apparent was the cost of abstraction in the non-monetary sense. Hermes sits between you and your tasks. That is the point of it, you delegate to it and it handles the details. But this means that when you want to make a small change, like adjusting the time a task runs, you go through Hermes to do it. You describe the change in natural language, Hermes interprets it and updates its internal configuration
Directly editing a crontab takes ten seconds and costs nothing. Doing the equivalent through Hermes costs credits and requires the system to correctly understand your intent. For the most part it does, but you are adding a layer of indirection to operations that did not previously need one.
The same applies to finding things. If Hermes has stored something in your Obsidian vault, you can search the vault directly, or you can ask Hermes where it put something. The latter is more convenient in the moment but creates a dependency. Over time you find yourself relying on Hermes to locate things it created on your behalf, rather than knowing where things are yourself. Of course, this again costs credits.
The Personal Assistant Analogy
A personal assistant is genuinely useful. They remember your preferences, handle routine tasks without being asked, and reduce the cognitive overhead of administration. But they come with ongoing costs. You pay a salary. You spend time briefing them. When they are unavailable, you may not know where things are or how things were set up because the knowledge lives with them rather than with you.
Hermes is the AI equivalent of this. The abstraction is real and it works, but it comes with the same trade-offs. There are upkeep costs — credits — that accrue even on days you do nothing. And there is a degree of opacity that is inherent to delegation. You hand things off and they get done, but the implementation details are no longer yours.
Conclusion
Hermes is a coherent product and the memory integration is genuinely interesting. But it is not the right fit for me. For the tasks I was running, the simpler approach is to write a few scripts in Claude, set them up in cron directly, and know exactly what is running and when. No ongoing credit consumption outside of when I actually want AI involved. No abstraction layer between me and a crontab edit.
Whether Hermes makes sense depends on how much you value not thinking about the details. If you want a true set-and-forget assistant and the credit cost is not a concern, it does what it promises. If you are comfortable with a bit of setup and prefer to stay in control of your own systems, the overhead probably is not worth it.