Updated
How much do AI agents spend rereading the same files?
There's no single honest dollar figure: it depends on your repo size, how often agents run, and your model's pricing. But the waste is structural, the same docs reread every session, for every agent, for every teammate. The measurable part is per-lookup: serving the one canonical section is about 60% fewer tokens than reading the top stale candidates.
Why the cost compounds (and is easy to miss)
A single reread looks cheap. The bill comes from repetition: an agent re-establishes context at the start of most sessions, you run several agents, and your teammates' agents do the same on the same repo. The same lookup gets paid for again and again, so a "small" per-session cost multiplies across sessions × agents × people. Bigger context windows don't help — you still pay per token, you can just load more of the same.
Where the 60% comes from
It's the difference between two behaviors on one doc lookup: reading the top few candidate files (including stale and duplicate copies) to figure out which is current, versus being handed the single canonical section that answers. The latter is roughly 60% fewer tokens because the agent reads one current passage instead of several near-duplicates. That's a per-lookup number, measured on real repos — not a promise about your total bill, which depends on how many lookups you do.
Get your real number
trovex keeps a savings dashboard that shows, on your own repo, what the agents would have read versus what they actually read — so you can replace the estimate above with your measured figure. It runs locally (SQLite + ONNX, no cloud or keys).
FAQ
How much does the rereading actually cost?
It depends on your repo, but the pattern is the same: the agent reads several candidate files every session to find the one current answer, and most of those tokens are discarded. On our own repo, serving one canonical doc instead cut a median of 69% of tokens per lookup at equal task-success, range 41 to 81%; we headline about 60%.
Is it cheaper to just let it reread?
Only until the bill arrives. Rereading is paid on every session, by every agent, for the same answer. Index once and serve the current doc per query, and a recurring cost becomes a one-time one.
Does this need a cloud service?
No. trovex runs locally with vectors in SQLite and an on-device embedder, no cloud or API keys, so your files never leave your machine.
Measure what you'd stop spending.
trovex is open source and in public beta. Install it and see the savings on your own repo.
uv tool install trovex
Open source (AGPL-3.0 core, MIT CLIs). Local-first — your docs never leave your machine.