A client asks for a page it has already fetched. The server replies with a status code and no new copy of the page. That small exchange can be enough to reuse stored content, provided the client has kept the right context.
For automated research, it is worth distinguishing the content saved by a cache from the conclusion an agent previously drew from it. The web protocol can help with the first. The second requires remembering what question was being answered.
A validator for a representation
HTTP conditional requests let a client send a validator such as an entity tag with its request. A successful validation of a cached GET can produce a 304 response instead of a new response body. The client uses its stored representation.
Imagine a documentation page describing an API’s input fields. A research tool saves the page and its validator. On the next fetch, a 304 can avoid downloading that representation again. If the task is to detect a documentation change, that is useful information.
Which version was stored?
HTTP caching rules for the Vary header address a further detail: selecting a stored response can depend on request fields. The URL alone need not identify the representation that is appropriate for a new request.
A research archive should therefore retain the request context that mattered. A page fetched in one language, or through an authenticated session, may not answer the same question as a later public request. Reusing a body without that context can save bandwidth while losing meaning.
The question can change first
Suppose the archived page said an integration required an API key. Yesterday’s question was how to authenticate. Today’s question is whether a particular account has a valid key. The document could be byte-for-byte unchanged and still be insufficient for the new question.
The practical record has two parts: the fetched document, with its time and request context, and the claim made from it. Refreshing a document can detect a changed source. Revisiting a claim can detect that the source was never enough to support the new use. Keeping those records separate makes both checks easier.
Published . Corrections