The model is not allowed to remember

Ask a chatbot a question and it answers from memory. Everything it says comes from patterns fixed at training time, months or years ago, with no way for you to tell which parts are solid and which it has confidently made up.

That is the actual problem with AI search. Not that models are bad at language, but that they are used in a way that makes their output impossible to check.

The rule

Foxhole Web has one rule that shapes everything else: the model does not answer from memory. Every question runs a live web search first, and the results of that search are the only material the model is given. The prompt says so explicitly, and it also says that if the results do not answer the question, the model should say so rather than fill the gap from training data.

It does not always obey perfectly. Models are not deterministic and instructions are not laws. But the ceiling on how wrong it can go is much lower when the material in front of it is fresh and cited.

What that buys you

Every claim carries a number, and every number points at the page it came from. When an answer matters, you click through and read the source yourself. The answer stops being something you either trust or do not, and becomes a starting point you can verify in about ten seconds.

It also means answers stay current. Ask about something that happened this morning and you get this morning's pages, not a model's recollection of the world as of its training cutoff.

What it costs

Speed, mostly. A chatbot answers instantly because it is reading its own weights. Foxhole Web has to run a real web search, wait for results, then read them. That is a few seconds of genuine work, which is why the answer streams in as it is written rather than appearing all at once.

It costs money too. Every single search is a paid search API call plus paid inference. There is no version of this that is free to run, which is a problem we will have to solve in public.

The honest caveat

None of this makes the answers true. A model can misread a source, or cite one that does not quite support the claim it is attached to. Sources themselves are often wrong.

What the citations give you is the ability to find out. That is a smaller promise than "trust our AI", and a much more useful one.