Ask a question, get an answer, close the tab. The next question starts from zero, paying full price for context the last one already bought.
Research done through a chat window is wasteful in a way that is easy to miss. The answer arrives, the tab closes, and the sources and the reasoning go with it. The next question, however adjacent to the last, starts from nothing and pays full price again for context that was already bought.
Hunty keeps it. One command finds the links, fetches them, synthesises a cited answer, and writes every source and the answer into a local store on my own machine. Later questions run against that store instead of the open web. The writing side and the reading side share the same tables and the same index, which is a boring detail that turns out to be the entire product.
It is boring and it is not obvious. The tool names the predecessors it was built away from, and one specific defect that set one of them back: a separate facts table that the answering side never actually read. A write-and-read mismatch does not fail loudly. It just never compounds. In the Hunty record, that mismatch stalled Senti for a month.
The defaults are the other reason I kept it. Synthesis is extractive by default with no model in the loop at all, and putting a model over untrusted web text is opt-in and sandboxed, because even a read-only agent reading a hostile page can pull local files into its answer. A run that finds no usable sources aborts rather than storing a confident answer with nothing behind it. I would rather have a tool that fails loudly than one that fabricates politely.