What is context decay?

Context decay is the gradual erosion of the brand context surrounding an AI system, while the output keeps arriving as if nothing changed. It happens between sessions (each chat starts from zero), between people (the best prompt lives in one person's notes), between revisions (a rule agreed in review is applied once and never written down), and over time (the corpus still contains positioning you have since moved on from). The output stays fluent throughout, which is what makes it hard to spot — the work does not look broken, it just quietly stops sounding like you.

Is this not just a prompting problem? Better prompts, better output.

Better prompts genuinely help, and they are the right first move. But a prompt raises the floor for one generation, in one session, for one person — and it decays the moment that session ends. If the fix has to be re-applied by a human every time, the failure mode is structural rather than instructional. The question worth asking is not whether your team can write a good brief; it is whether the good brief survives the person who wrote it going on holiday.

Does this eliminate off-brand output?

No, and we would not claim it does. Generation is probabilistic and judgement is human — some drift is inherent to both. What changes is that drift becomes detectable and correctable rather than invisible: the rules are in one place instead of scattered across notes, the evidence behind a generation stays attached to it, and when something does come out wrong you can see what it drew on rather than guessing. That inspection happens inside your own CrawlQ tenant — it is an editorial review capability, not an independent third-party audit, and the limits on independent verification of the full chain are set out in our Article 50 piece. Detectable is a much lower bar than eliminated, and it is the honest one.

How is this different from fine-tuning a model on our content?

Fine-tuning bakes tone and vocabulary into the weights, which is genuinely useful for voice consistency. Two things it does not give you: it is slower and costlier to update when positioning shifts — less so with lighter adapter methods such as LoRA, but the update cycle is still a retraining step rather than an edit — and it is opaque, in that you cannot ask a fine-tuned model to show which source supported a specific claim in a specific asset. Governed context is faster to update and stays inspectable. They are not mutually exclusive; they solve different halves of the problem.

What does 'memory infrastructure' actually mean here?

It means brand context is held as governed infrastructure rather than as text somebody re-types into a prompt box: the corpus, the voice rules, the evidence, and the record of what was generated from them, kept outside any single session so every generation draws on the same current source. To be precise about the wording: this is governed context storage that you control and maintain, not a claim that our system silently retains everything it has ever seen. The verifiable end of it — cryptographic receipts for governed decisions, anchored to a public transparency log — is shipped today; further provenance layers such as embedded content credentials are roadmap rather than shipped, and the full ledger of which is which is set out in our piece on Article 50 and replayable records. The short version: what an organisation says is one problem, and being able to show what its AI did is a different one.

Product field note · Current as of 27 July 2026

Why AI forgets your brand.

The output was good in March. By June the same team, using the same tool, is producing work that is fluent, competent, and somehow not quite you. The usual diagnosis is that people got lazy with their prompts. The more accurate one is that the brand context around the system was never held anywhere durable — so it decayed, quietly, while the output kept arriving as though nothing had changed.

Context decay — the gradual erosion of the brand context surrounding an AI system, while its output continues to arrive with undiminished confidence. It is hard to spot precisely because nothing looks broken.

Between sessions
Every new chat starts from zero. The brand context that made yesterday's output good is not present today unless somebody re-types it.
The tell
The same brief produces noticeably different work on Tuesday than it did on Monday.
Between people
The person with the best prompt has the best output. That prompt lives in their notes, not in the system, so quality tracks the individual rather than the team.
The tell
Output quality drops when one particular colleague is on leave.
Between revisions
A rule is agreed in a review — never say that, always cite this — and applied to the asset in hand. The rule itself is not written anywhere the next generation can read.
The tell
The same correction gets made in review after review.
Over time
Positioning evolves. The corpus the model draws on still contains last year's claims, with nothing marking which version is current.
The tell
Retired messaging resurfaces in new drafts months after the reposition.

This is not a model problem

It is tempting to read drifting output as a capability ceiling — the model is not good enough, so the work is not good enough. That reading is usually wrong, and it is expensive, because it sends teams shopping for a better model when the actual failure sits somewhere else entirely.

A capable model with thin context produces confident, fluent, generic work. That is the signature failure of context decay, and it looks nothing like a model failing. There is no error, no refusal, no obviously wrong answer — just a slow slide toward competent output that any competitor could have published. The fluency is what disguises the problem. If the work came out visibly broken, somebody would have escalated it in week one.

The diagnostic question is simple. When output quality drops, ask whether the person prompting had the full brand context in front of them. If the honest answer is no, they were working from memory and a half-remembered doc, the model was never the constraint.

Why better prompting stalls

Better prompts are the right first move and they genuinely work. The problem is that they work in exactly the scope they are typed into: one generation, one session, one person. The context window is not storage. Whatever brand context you paste in competes for room with the actual task, and it is gone when the session closes.

Which produces the pattern most teams eventually recognise: the output is best when your sharpest person is doing the prompting, and it degrades when they are not. That is not a training gap. It is what happens when the important context lives in one person's notes rather than in the system everyone shares.

Four fixes, and where each one stops

A longer prompt
What it does
Raises the floor for one generation, in one session, for one person.
Where it stops
It decays the moment the session ends, and it competes for the same context window as the actual task.
A prompt library
What it does
Makes the good prompt shareable, so quality stops tracking one individual.
Where it stops
It is a snapshot. Nothing updates it when positioning changes, and nothing records which version produced which asset.
Fine-tuning
What it does
Bakes tone and vocabulary into weights, so the voice holds without re-prompting.
Where it stops
It is slow to update, opaque to audit, and answers nothing about why a specific asset came out the way it did.
Governed brand context
What it does
Holds the corpus, the rules, and the evidence outside any single session, so every generation draws on the same current source.
Where it stops
It needs the context to be maintained as a real asset. Software can hold it and show its state; it cannot decide your positioning for you.

What a structural fix looks like

If the failure is that context does not survive the session, the fix is to stop keeping it in sessions. That means holding the corpus, the voice rules, and the evidence as governed infrastructure — outside any single chat, readable by every generation, and updated in one place when positioning moves. This is what we mean by brand context as infrastructure rather than as text somebody re-types. The shape of that is described on the knowledge graph page, and the mechanics on how GraQle works.

Three properties matter, and it is worth being specific about what each one buys. One current source — when positioning changes you update the context, not fourteen prompt docs, so retired messaging stops resurfacing. Rules that outlive the review — a correction agreed once is written where the next generation reads it, instead of being re-made in review after review. And evidence that stays attached — the sources behind a generation travel with the asset, so a claim in a draft can be traced rather than defended from memory.

None of this makes drift go away. Generation is probabilistic, judgement is human, and some slide is inherent to both. The honest claim is narrower and more useful: drift becomes detectable and correctable rather than invisible. You find out in the draft instead of in the quarterly review.

The same problem, in its compliance form

There is a second reason to care, and from August 2026 it stops being optional for anyone publishing AI-generated content into the EU. Article 50 of the EU AI Act requires that such content be marked and disclosed (not legal advice — scope depends on your content, your systems and your jurisdiction, and that is a judgement for your counsel) — and behind that sits a harder question: when somebody asks how a specific asset was produced, can you answer with a record rather than a recollection?

That is the same decay, wearing different clothes. A team that cannot reconstruct why its AI wrote what it wrote has a brand problem on Monday and an evidence problem when the regulator writes. The compliance half of this argument — including which parts of the provenance stack are shipped today and which are still roadmap — is set out in Article 50: why compliance requires replayable memory. For the obligations mapped to specific mechanisms, with the limits of each stated, see how brand memory changes your EU AI Act exposure. The controls behind those claims are listed on our trust page, and none of this is legal advice.

Where to start

The cheapest useful move is a diagnosis, not a purchase. Take a piece of output from this month that did not feel right, and ask what the system knew when it produced it. Not what your team knows — what was actually in front of the model. In most cases the gap between those two is the whole story, and it is measurable before anyone buys anything.

If the gap is large, the fix is structural rather than a matter of trying harder, and the good news is that structural fixes hold. Context that lives in infrastructure does not need to be remembered by whoever happens to be at the keyboard on a Tuesday in November.

Questions we get asked

What is context decay?: Context decay is the gradual erosion of the brand context surrounding an AI system, while the output keeps arriving as if nothing changed. It happens between sessions (each chat starts from zero), between people (the best prompt lives in one person's notes), between revisions (a rule agreed in review is applied once and never written down), and over time (the corpus still contains positioning you have since moved on from). The output stays fluent throughout, which is what makes it hard to spot — the work does not look broken, it just quietly stops sounding like you.
Is this not just a prompting problem? Better prompts, better output.: Better prompts genuinely help, and they are the right first move. But a prompt raises the floor for one generation, in one session, for one person — and it decays the moment that session ends. If the fix has to be re-applied by a human every time, the failure mode is structural rather than instructional. The question worth asking is not whether your team can write a good brief; it is whether the good brief survives the person who wrote it going on holiday.
Does this eliminate off-brand output?: No, and we would not claim it does. Generation is probabilistic and judgement is human — some drift is inherent to both. What changes is that drift becomes detectable and correctable rather than invisible: the rules are in one place instead of scattered across notes, the evidence behind a generation stays attached to it, and when something does come out wrong you can see what it drew on rather than guessing. That inspection happens inside your own CrawlQ tenant — it is an editorial review capability, not an independent third-party audit, and the limits on independent verification of the full chain are set out in our Article 50 piece. Detectable is a much lower bar than eliminated, and it is the honest one.
How is this different from fine-tuning a model on our content?: Fine-tuning bakes tone and vocabulary into the weights, which is genuinely useful for voice consistency. Two things it does not give you: it is slower and costlier to update when positioning shifts — less so with lighter adapter methods such as LoRA, but the update cycle is still a retraining step rather than an edit — and it is opaque, in that you cannot ask a fine-tuned model to show which source supported a specific claim in a specific asset. Governed context is faster to update and stays inspectable. They are not mutually exclusive; they solve different halves of the problem.
What does 'memory infrastructure' actually mean here?: It means brand context is held as governed infrastructure rather than as text somebody re-types into a prompt box: the corpus, the voice rules, the evidence, and the record of what was generated from them, kept outside any single session so every generation draws on the same current source. To be precise about the wording: this is governed context storage that you control and maintain, not a claim that our system silently retains everything it has ever seen. The verifiable end of it — cryptographic receipts for governed decisions, anchored to a public transparency log — is shipped today; further provenance layers such as embedded content credentials are roadmap rather than shipped, and the full ledger of which is which is set out in our piece on Article 50 and replayable records. The short version: what an organisation says is one problem, and being able to show what its AI did is a different one.

See what governed brand context looks like.

The brand governance framework page walks through the scoring, the grounding, and the audit trail — with the limits of each stated plainly.

Brand governance framework How GraQle works