Every agent memory system I have read stores what was said and when. Almost none store where. The chunk knows its text, its embedding, a created_at, maybe a session id. It does not know that you learned it on a Wikipedia article, or typed it into a vendor onboarding form, or that it came out of a ChatGPT conversation you had on a specific page.
That missing axis costs you more than it looks like it should. Without it you cannot answer "what do I already know about the thing in front of me," you cannot forget one source without hunting chunk ids, and you cannot tell a user which of your stored facts came from their bank's website — which, the moment you have a browser extension, is the question that decides whether they trust you at all.
I spent two days adding that axis to Vodou's memory store. Here is what it actually took, including the four defects that every unit test in the repo said did not exist.
What a page axis is
Three columns on the memory chunk: source_url, source_host, source_ref. Then a match endpoint that, given a URL, returns what memory holds about it in four tiers.
