Context decay
Your context does not decay at one speed
Hand an organization's knowledge to an AI and part of it is stale by next week, part will still be true in ten years. The mistake is treating it as one substance with one shelf life. It is layered. Each layer has its own half-life.
(noun)/ˈkɒn.tekst dɪˈkeɪ/
The kinds of context inside one company decay at very different speeds. Some rot in weeks, some hold for decades.
Treat all your context as one thing and the fast layers rot and mislead your AI, while the slow layers, the ones that make you you, never get captured at all.
Similar institutional memory, know-how, silent knowledge, tacit knowledge, tribal knowledge
One archive, six shelf lives
When people picture "capturing our context" they picture a document: write it down once, hand it to the AI, done. But a company does not hold one kind of knowledge. It holds this week's messaging and its reason for existing in the same breath, and those two things age nothing alike.
Some context is perishable. Your current roadmap, your live campaign, the price you quote today. Some is close to permanent. Your values, why the company exists, the taste that runs through everything you make. Between them sits everything else, each with its own rate of going out of date.
Slow layers make it sound like you.
Reachable is not the same as true
Dark Context asks two questions about what your organization knows. Can it be reached at all? The knowledge nobody ever wrote down is dark context. And is it still true? Captured knowledge nobody keeps fresh is gray context, the slow slide from current to quietly wrong.
Context decay sharpens that second question. A fast startup's context goes gray in weeks, a stable institution's in years. But decay is not uniform even inside one company: the same organization holds context on every layer at once, and each layer ages at its own pace.
Fast at the rim, still at the core
The outer layers churn; the core holds. Speed of motion = speed of decay.
From the fast rim to the still core
What each layer holds, and how quickly it goes out of date.
- VoiceweeksCurrent tone, phrasing, what you are saying this week, the live campaign.
- OperationsmonthsPrices, offers, live projects, the roadmap, who is on what this quarter.
- Systemsa couple of yearsTools, stack, processes, the org chart, roles, vendors.
- Strategyseveral yearsPositioning, decisions and their reasoning, policies. Breaks the moment an assumption behind a decision breaks.
- Culturea decadeHow you actually decide, your taste, your standards, the unwritten rules. Mostly never written down.
- Corenear-permanentValues, mission, why you exist, the founder DNA the whole company defaults to.
The rim rots first
The outer layers are where gray context is made. They change constantly, so they need constant upkeep, and an archived copy is stale almost as soon as it is written. This is where an AI turns confident, specific, and quietly out of date.
A customer asks your AI what onboarding costs. It answers instantly, confidently, with the exact figure from a deck uploaded in March. The price changed in May. Nobody lied. The archive did.
The fix is not a better archive. It is capture on demand. Ask "what is the roadmap today" and get the current voice, not the stale record. The fast layers want a live answer, not a filed one.
The still layers are the asset
The inner layers barely move. Values, culture, why you exist: capture them once and they hold for years. They are also the layers that make your output yours. A general model knows almost everything in general and almost nothing about your core, so when it fills that gap it guesses, and the guess is slop.
Bring the slow layers out of the dark and the AI stops sounding generic and starts sounding like the company. This is the anti-slop asset. It is also the cheapest to maintain, because it hardly changes.
The slow layers are cheap to keep and expensive to lose.
Sweep each layer at its own speed
Gray context said sweep at the speed of your ship. Context decay makes that precise: sweep each layer at its own speed. You re-check Voice weekly and the Core almost never.
Put every layer on one calendar and you lose either way. Refresh it all constantly and you drown maintaining a core that did not need touching. Refresh it all rarely and the edges rot while you tend the middle. Match the cadence to the layer.
Start at the core, not the rim
Capture the slow layers first: core, culture, strategy. They are durable, high-leverage, and the one thing a general AI can never guess. Do them once and they pay off for years.
Then put the fast layers on a refresh cadence, captured on demand rather than filed once and forgotten. The core makes the AI you. The rim keeps it current. You need both, at different speeds.
Bring your core out of the dark
Context decay is one lens in a larger project about the knowledge organizations never wrote down, and what it takes to hand it to an AI without losing what makes you you.
darkcontext:~$which layers are going gray fastest?
A short call, not a sales pitch. Bring one process whose context you want to keep current, and we'll work out which layers to sweep and how often — and whether it's worth it.
book a calldarkcontext:~$not ready to talk?
Leave an email instead. No pitch, just the occasional note as we map where dark context shows up and what helps bring it to the surface.