A system that never forgets will eventually be confidently wrong.

Ask any vendor how their system handles knowledge that has become false. Most will talk about re-indexing, which answers a different question. Re-indexing keeps the store current with the documents. It does nothing about a claim that was true in March and is not true now, because no document was written to say so.

This is the ordinary case, not the edge case. A competitor exits a segment. A plant shuts. A buying manager retires and the new one has different priorities. A price holds for a season. Nobody files an update. The claim simply stops being true, silently.

Half-lives by knowledge type

Different facts age at different rates, and treating them uniformly is what produces absurd behaviour. A quoted price has a half-life measured in weeks. A plant location has one measured in years. A decision maker's stated preference sits somewhere in between, and a structural fact about how an industry buys may barely decay at all.

So half-lives are set per knowledge type, agreed at build time, and reviewed in the quarterly. Confidence decays along that curve, and contradicting evidence does not average against the existing claim, it collapses it. One credible contradiction should beat six months of accumulated agreement, because the world changing is more likely than six months of observation being wrong.

Stale knowledge should be held out of decisions, not quietly reused because nothing has replaced it yet.

What it looks like in practice

A claim that has aged past its threshold is marked stale. It stops informing decisions and it appears in a queue for re-evidencing. If the evidence comes back, confidence resets. If it does not, the claim is retired and the system is explicitly less certain than it was, which is the correct state to be in.

People find this reassuring in a way that surprises first-time buyers. A system that admits the limits of what it currently knows gets trusted with more, not less.