How do you know if the answer in your docs is still true?
Most search tools return the passage that best matches your question. Whether it is still true is not something they model. Here is what does, and how to tell the difference.
You don't — unless the system stores when each fact was last confirmed. Most search tools, and every AI assistant built on document retrieval, return the passage that best matches your question. Whether that passage is current is not something they model, so a fact that changed six months ago comes back with the same confidence as one confirmed yesterday. Headknot stores facts as dated claims rather than as text to summarise, so an answer carries the date it was last true and the sources behind it.
Why AI search makes stale documentation worse, not better
Document search at least shows you a file's age, which leaves the judgement to you: you see a page last edited in March and decide for yourself whether to trust it. An assistant that summarises removes that signal. The answer arrives fluent, confident and undated, and the one cue you were using to catch staleness is gone. This is the specific failure people mean when they say an AI tool "made things up" about their own company. Usually it invented nothing — it faithfully reported something that used to be true. The information was real. The timing was not, and nothing in the interface told you which.
What "dated claims" means in practice
A claim is a single fact with a subject, a value and a time: Pro costs $12 a month, valid from 4 June 2026. When the price changes, the old claim is not overwritten — it is closed, and a new one opens. The record of both survives. That is what makes three otherwise impossible questions answerable: what is true now, what was true on a given date, and what changed between them. Document retrieval cannot answer the second or third at all, because a document has one state — its current text — and no memory of what it said before.
Which tools tell you when a fact changed?
Notion AI, Slack AI, Glean and Guru all search your content and cite their sources, and all four are good at that. None of them models time. They return the passage that ranks highest and leave you to work out whether it is current. Guru comes closest by asking a human to periodically verify a card, which works until nobody does it. The distinction is not about quality — it is structural. A system that stores documents can tell you a document is old. A system that stores dated claims can tell you a fact changed, when it changed, and what it was before.
How do you find out what a value used to be?
You ask for the date you care about. "What was our Pro price in April?" returns the claim that was valid in April, not the one valid today, along with the sources that supported it at the time. The same mechanism answers "what changed this quarter" and "who owned this before". None of it requires anyone to have written a changelog, because the history is a by-product of how facts are stored rather than a document somebody has to remember to maintain — which is the part that always stops happening about six weeks in.
How do you stop documentation going stale in the first place?
You mostly don't, and planning around that is more honest than another review process. Documentation goes stale because the decision moves before the doc does — the change happens in a thread, a ticket or a call, and updating the page is a separate act of work that competes with everything else. The realistic approach is to read all of those sources rather than only the tidy one, and to treat the most recent statement of a fact as the current one. That turns staleness from something a team has to prevent into something the system resolves at the moment a question is asked.
Headknot connects to Slack, Notion, Google Drive, Jira, Confluence, GitHub and Linear, and answers questions with the sources behind each fact and the date it was last true. Free to start, no card — 200,000 words to get going.
Start free — no card