Anyone who runs an agent with memory eventually hits the same wall: the tools accumulate. And at some point you ask yourself — shouldn't I throw some of them out?
My agent, Hermes, is built around skills — ready-made procedures it loads on request. They can be unlimited, unlike memory. But everything has a price. And this is where it gets interesting.
A skill isn't a note — it's a card in a catalogue
The key thing to understand: a skill works in two stages.
First, the agent only sees each skill's name and first line of its description — all of it is baked into the system prompt on every message you send. It's like a library catalogue: you don't read every book, you only see the spines on the shelf.
Then, once your question matches a skill by meaning, the agent loads its full text.
Those "spines" are what you pay for constantly. A hundred, two hundred skills — and each time, on every request, you carry that catalogue with you. Even if none of them is needed right now.
Why "just delete it" is a bad idea
The first instinct is to remove everything old and unused. Logical, but dangerous. A skill is hard-won experience: how to do a thing, what the pitfalls are, which command works and which doesn't. Throwing it away means losing accumulated knowledge.
So — don't delete. But keeping two hundred skills in the catalogue is a permanent tax on every request and an increasingly imprecise search: the more similar skills there are, the more often the agent picks the wrong one.
The archive is the third path
The archive is the middle ground. A skill is taken off the "shelf" (it stops costing you as a catalogue spine), but it's not destroyed. It stays in a special folder, and you can bring it back in one click.
In my Hermes, an automatic curator does this — a background task that looks at the library weekly and:
- merges the similar — three skills about the same thing collapse into one, with their experience carried into the shared skill;
- archives the unused — if a skill hasn't been opened in a month or so;
- never deletes — everything lands in
.archive/and can be restored.
I had one of these passes recently: three token-compliance skills merged into a single umbrella, six GitHub skills went to the archive (an up-to-date one already covered them), and a couple of macOS-only ones were pruned because my server runs Linux — they physically can't run here.
Why it's not about "money"
It's easy to make a mistake here and decide archiving exists so you don't pay for expensive prompts. It doesn't.
Running a skill costs the same — whether it's active or just restored from the archive. The saving isn't there. The saving is that the constant background from hundreds of unneeded descriptions disappears, and the catalogue becomes more precise: less clutter means fewer wrong choices.
Frame it like this: an agent's memory is limited, skills are not. But skills have their own hidden price — a place in that catalogue you carry with you on every turn. The archive lets you keep the catalogue clean without losing anything.
The takeaway
Skills are knowledge without a volume limit. The archive is a way to put that knowledge in order without risking it.
For me it's a simple formula: a skill I need — active. A skill I once needed — archived, just in case. And a skill I'll never need again, I still won't delete — you never know when it'll come in handy tomorrow.
