Most AI workflows treat sources as garnish. The model produces an answer, the user copies it, and the citations stay buried in the chat sidebar. That's fine for casual use. It's dangerous when the work has consequences.
Three failure modes
We've seen the same three failures over and over.
The phantom citation. The model confidently cites a paper that doesn't exist, or misattributes a claim. You publish, someone checks, and the trust deficit lands on you. not the model.
The lost provenance. The citation was real, but you can't find it three months later. The answer survives; the receipt doesn't. When a stakeholder asks “where did this come from?” you're back to the search bar.
The stale source. The citation was real and verifiable, but the underlying page changed. The claim you made is no longer supported by the link you saved.
What changes
Treating sources as first-class objects fixes all three. When you save a link, Shelvia records the URL, the title, the fetch date, and the project the source helped. When you save a passage from a chat or document, the original text stays attached so you can verify it later. When sources connect to decisions, the reasoning behind a call stays defensible.
It's not a feature. It's a posture: don't separate the answer from the receipt.
What this gets you in practice
Six months from now, when a colleague asks “why did we make that pricing call?”, you can answer with the conversation, the competitor research, and the decision pinned at the time. When you build on top of yesterday's research, the synthesis isn't a leap of faith. it's a continuation. When you onboard a new teammate, the trail of reasoning is readable.
That's the kind of memory AI work deserves.
