Why does system-prompt adherence decay over a long conversation?

## What drift actually looks like Instruction drift is the gradual erosion of system-prompt adherence as a conversation gets longer. Nothing dramatic happens at any single turn. A tutor told to reply only in Spanish answers in Spanish for twenty-five turns, then starts glossing a grammar point in English, then answers the next question mostly in English. A support assistant told to end every reply with a ticket reference does so reliably early and omits it by turn forty. A model told to keep answers under three sentences slowly returns to its default essay length. Note that the system prompt is almost always still there — chat APIs are stateless, so the client resends the whole conversation, system block included, on every request. Presence is not the problem; **salience** is. ## Why the rule loses ground **Distance and competition.** The system prompt is the oldest text in the request. Early in a chat it is a large share of everything the model is conditioned on. By turn fifty it may be one percent of the window, sitting far from the position where generation begins, competing with thousands of tokens of newer, more concrete, more topically relevant material. Attention is finite and is spread across the whole sequence; a short abstract rule is easy to out-shout with a long specific discussion. This is the same phenomenon people describe as context rot: quality of instruction-following falls as the occupied window grows, even well inside the nominal limit. **Self-imitation.** This is the effect people underestimate. Once the model produces one reply that breaks the rule, that reply is appended to the history as an assistant turn and becomes part of the conditioning for every subsequent turn. In-context examples are the strongest steering signal available to a language model, and the transcript is now an example set demonstrating the wrong behaviour. So drift is not independent per turn — it is a ratchet. Recovery rarely happens on its own; the conversation has to be pushed back. **Instruction load and salience ranking.** Rules are not equally sticky. A prompt with fifteen standing constraints does not decay uniformly: constraints that are reinforced by the content of the conversation survive longer, and constraints the dialogue never touches fade first. Tone, persona and formatting rules typically slip earliest because nothing in the exchange re-evidences them; a rule that the user's own messages keep implicitly reinforcing ("in grams, please") survives longer. Practically, ten rules do not get ten times the adherence of one — they dilute each other. ## Turns or tokens? Turn count is the convenient proxy, but token distance and content volume are the better predictors. Forty short turns totalling three thousand tokens drift far less than forty turns totalling ninety thousand tokens with pasted documents and long tool outputs in between. Two other factors matter: how much of the recent window is model-generated (self-imitation pressure), and whether anything has re-stated the constraint recently. Two conversations at the same turn number can be in completely different adherence regimes. ## Which failures matter Separate **persona and tone slip** — the assistant stops sounding like the character, gets more verbose, drops a stylistic convention — from **hard rule violation** — it discloses something it was told never to disclose, gives advice it was told never to give, or breaks an output format a downstream parser depends on. Both are drift; only one is usually an incident. Which of the two your product cares about determines how much you spend fighting drift and where you enforce the rule. ## How you notice You will not see this in a single-turn test set, because every case there starts at turn one where adherence is at its best. Drift is only visible when you replay long conversations and check the same constraint at several depths — for example asserting it at turn five, turn twenty and turn fifty — and look at the adherence curve rather than a pass/fail. In production, the same idea shows up as adherence rate bucketed by conversation length or by tokens of history; if the metric falls off a cliff past some depth, that is your practical ceiling for unassisted adherence. ## What it is not Drift is not the model deciding the user outranks the system prompt, and it is not the system prompt being stripped or expiring. It is also not randomness alone: temperature adds variance, but the trend across a long chat is directional, driven by dilution and by the model imitating its own recent output. Diagnosing it as "the model ignored my instruction" leads people to write a longer, angrier system prompt, which usually makes things worse — more instructions means more dilution.