There is a comfortable fiction in how organizations think about context. The fiction is that context is a thing you can collect - that if you gather enough documents, index enough knowledge, and connect enough systems, you will have the context, and it will sit there, accurate and available, waiting to be used. Build the corpus once and draw on it forever.
Context does not work like that. Context has a half-life. It begins decaying the instant it is created, and a knowledge base assembled and left alone is not a stable asset. It is a slowly spoiling one, growing more confidently wrong with every quarter no one tends it.
Why context decays
Decay is not a flaw in how you captured the context. It is a property of what context is. Context describes a situation, and situations change. Consider what happens to a single captured fact over time.
A policy was current when it was written into the knowledge base, and then it was superseded - but the old version is still indexed, still retrievable, still phrased with total confidence. An ownership mapping was accurate, and then the person left and the team reorganized, and now it points to no one. A precedent was relevant, and then the circumstances that made it relevant stopped applying, but nothing flagged it as stale. Each of these facts was true. None of them is current. And a system that cannot tell the difference will serve the stale version with exactly the same fluent confidence as the fresh one.
Wrong context is more dangerous than missing context. Missing context produces an “I don’t know,” which a person can act on cautiously. Stale context produces a confident, well-formed answer that is quietly false, which a person acts on as though it were true.
This is the asymmetry that makes context decay so corrosive in AI-native systems. The model has no native sense of freshness. It treats a fact captured this morning and a fact captured three years ago as equally available, equally authoritative. Without an explicit model of decay, the system’s confidence is uncorrelated with its currency - and confidence uncorrelated with truth is the precise definition of a system you cannot trust.
The fix is not “refresh more often”
The instinct, once you see the decay, is to schedule re-indexing. Crawl the sources again, rebuild the corpus, refresh the knowledge base on a cadence. This helps at the margin and misses the central problem, because the most valuable context was never in a source you can re-crawl.
The reasoning behind a decision, the assumptions in force when a choice was made, the obligations that a particular commitment created - this context exists only at the moment the decision is made, and only in the heads of the people making it. There is no document to re-crawl. If you do not capture it then, it is gone, and no amount of later re-indexing will recover it, because it was never written down anywhere a crawler could reach.
This is why the discipline that actually addresses decay is capture-at-decision: recording context at the moment it is generated and still true, attached to the decision that produced it, with the provenance and timestamp that let a later system reason about its freshness. You do not fight decay by refreshing stale context faster. You fight it by capturing fresh context at the only moment it exists in full.
Decay as a first-class property
The deeper shift is to stop treating freshness as metadata and start treating it as a first-class property of every piece of context the system holds. Every fact should carry not just its content and its source but its age, its expected half-life, and the conditions that would invalidate it. A policy fact and a phone number decay at very different rates; a system that models both as “true until deleted” will be wrong about both, just on different schedules.
This is the part of the Context Intelligence Infrastructure that the model-centric view never sees. The infrastructure’s job is not only to know things. It is to know how long it has known them, how confident it should be given that age, and when a piece of context has decayed past the point where it should be served at all. A context layer without a decay model is a library that keeps reshelving retracted books next to current ones and trusts you to know which is which.
The closing thought
Treat context as a static asset and you will build a system that grows more confidently wrong over time, because the decay is invisible right up until it produces a costly, fluent, false answer. Treat context as something with a half-life - captured when it is true, stamped with its age, retired when it decays - and you build a system whose confidence actually tracks reality.
Context is not something you have. It is something you keep capturing, because the moment you stop, it starts becoming a lie that still sounds like the truth.