How much memory an AI agent really uses: measurements

0.36 GB idle, 0.81 GB with a browser, 0.85 GB on a parallel build. What that means for picking a plan, and how to measure it yourself.

We run agents on our own hardware and measure what they actually consume. There is almost no public data on this — the usual advice is "give it 4 GB to be safe". Here is what we found.

What Hermes actually uses

Running — Memory

Agent idle, no heavy tools — 0.36 GB

Plus Chromium, light page — 0.67 GB

Plus Chromium, heavy page — 0.69 GB, peak 0.81 GB

Parallel project build in the terminal — peak 0.85 GB

The gap between idle and peak is the whole story. At rest the agent is about a third of a gigabyte; one open browser page triples it. Plan for the peak, not the average — the OOM killer is not interested in your average.

What that means for a plan

An agent with a browser fits in one gigabyte, with roughly 190 MB to spare. That is enough while tasks arrive one at a time. Add a schedule that sweeps a dozen sources at once and the headroom is gone.

So on our side, personas that only write text sit on one gigabyte, and the ones that fetch from many sources on a schedule sit on two. Not for appearances — the first genuinely fit and the second genuinely do not.

What we do not know

Every number above comes from a single pod. We have not measured the spread between agents or between tasks; that needs real load, not a bench. A median of peaks across many pods is a more honest number than the peak of one, and when we have it we will publish it — including if it is inconvenient.

Measuring it yourself

Do not trust docker stats as your only source: it shows an instantaneous value and quietly misses short peaks. Sample rss every few seconds and take the maximum over the task window — the difference from the eyeball figure can be twofold.