rag gave my agents better recall, and they still drifted. retrieval finds a doc that looks relevant. it can't tell you which one is current, and it never writes back what the agent just figured out.…

trovex saved my agents 340,784 tokens (74%) in their first 2 days on a brand-new project. 91 doc-lookups, ~3,744 saved each. my run, not a universal number. i quote ~60% in public. open source,…

one slice i measured on my fleet: ~60% of a doc lookup's tokens went to re-reading the repo to guess which doc is current. a better model reasons better over that guess. it doesn't stop the guessing.

your agent isn't dumb for citing the stale runbook. retrieval ranks "most relevant," and a dead doc scores as high as the live one. relevance isn't currency. mark which doc is canonical, or it keeps…

we run agent fleets in prod. the bill climbs for one dumb reason: every session the agent rereads the repo to find which doc is current. one rollback lookup ~720 tokens, ~440 of them burned on files…