Main idea

A memory hook runs after every assistant response. At twenty exchanges a day, the spend stays under a dollar a month, and the processing takes one to two seconds — which assumes a small, fast model was chosen for the extraction task, not because it was good enough, but because the cadence demanded it.

What that level buys isn't a saving: it is the disappearance of a question. An automation costing ten times more wouldn't be unaffordable, it would be debatable — you would wonder whether to enable it on this project, disable it during debugging sessions, restrict it to important exchanges. Each of those questions comes up every time, and it is that repetition that wears you down, long before the bill does.

The useful threshold is therefore not "worth it" but "negligible". Below it, the setup becomes infrastructure you forget; above it, it stays a tool you decide about, and a tool you decide about at every exchange ends up being switched off.

Layer added by "Why You Should Audit Your Skills: Master the Hidden Token Limits" (2026-03-25). The same rule serves to decide where to start when several setups are already running. A skill launched ten times a week and a skill launched once a quarter don't call for the same care: it is frequency that designates the one that must be frugal, even before you know which is heaviest per run. And when the resource consumed isn't a bill but a ceiling — Claude Code's five-hour rolling window — the threshold of negligibility no longer reads in euros per month but in runs remaining before you're locked out.

Why it matters

This gives a design criterion for anything running in the background: size it not on the available budget but on the threshold beyond which a user would start wondering again whether to let it run.

It also justifies technical choices that would otherwise look timid — a small model, minimal processing, a thin extraction — because it is the cadence that constrains, not the difficulty of the task.

Nuances and limits

Cost per run isn't the only trade-off: one second of waiting per turn, acceptable on its own, becomes noticeable over a full day of continuous work.

And the threshold of negligibility is relative to who pays. What is painless for individual use becomes a budget line as soon as you multiply it by a team.

Open questions

• What do you do with an automation whose cost becomes debatable as it proves useful and you ask more of it?