flmnt → economics
Analysts forecast that AI coding costs will outgrow developer salaries, and a named driver is bloated context windows — history re-sent on every call. Budget-bounded retrieval is how we take that cost off the table: your agent gets the part of your history that matters for the task, so the cost of remembering stays flat as the project grows.
How much that's worth depends on one thing: how much history your project has accumulated. Here's how to tell which case you're in.
Scope
Which one applies to you depends on one number: how much project history you have accumulated. Not lines of code, not team size — the size of the record an agent would otherwise need to read.
Regime 1 · up to 50,000 tokens
difference is a wash
Below 50,000 tokens of accumulated history, the dollar difference is a wash. What holds at any size is currency: your agents work from what you decided, and the record survives every session.
Regime 2 · 50,000 up to the context window
a real, modest saving
Once the history is larger than what you'd want to send on every call, you start paying repeatedly to re-send things the agent doesn't need for the task in front of it. You stop paying to re-send your project's history, and the cost of remembering stays flat as the project grows.
Regime 3 · past the context window
a different question
When your history outgrows the context window, pasting it in stops being an option at any price. This is no longer an efficiency argument — it is the point where the alternative approach stops working. Our own project's record is around 645,000 tokens, roughly three times a 200,000-token window.
The claim
Re-sending the whole record is cheap and the agent is accurate. At this size there is nothing to fix.
Illustrative shape, not measured values. The measured figures are above, and on the benchmark with the method.
Two things happen as a project accumulates history, and they pull in opposite directions. The cost of re-sending the record climbs — slowly at first, then steeply. At the same time the agent gets less accurate, because a fuller context is a worse context: it contradicts decisions you already changed and repeats fixes that already failed. Budget-bounded retrieval starts higher, because retrieval has a floor cost that re-sending a short record doesn't have. The two cost lines cross at 50,000 tokens of history; past that point one keeps climbing and the other doesn't, and past the context window the record cannot be sent at all.
Method
SourcesThe forecast that AI coding costs will surpass average developer salaries, and bloated context windows as a driver, are Gartner's. The finding that committed context files tend to reduce task success while adding measurable inference cost comes from published research on instruction adherence. Both are linked on the benchmark page.
Sizing it for your project
Your record's size decides the argument. We'd rather run it against your project than a generic one.