Context Drift

Context drift refers to the gradual accumulation of noise, outdated facts, or incorrectly inferred content within the context window in multi-turn interactions or long-running AI systems, causing model output to systematically deviate from the original intent.

Addressing context drift requires engineering approaches, such as context versioning, periodic re-retrieval, explicit state snapshots and diff comparison, and regression testing based on evidence lineage.

Context drift differs from model hallucination: hallucination is a factual error within a single generation, whereas drift is a deviation that accumulates gradually across sessions or long reasoning chains.

Context drift refers to the phenomenon in which an agent gradually deviates from its original task scope or project boundaries during a long conversation.

Explicitly naming the project is regarded as the lowest-cost means of correcting context drift, because it provides an observable, reviewable boundary signal at every turn.

In multi-agent production systems, the first problem to appear is context drift, followed by retry amplification, and third is confident wrong answers (local models state wrong conclusions in the same format and tone, with no built-in uncertainty signal).

Structured data passing is the core means of mitigating context drift, and agents should pass typed fields (record ID, entity ID, timestamp, status code, confidence) between each other rather than natural language summaries.

Contextual drift refers to the phenomenon where a model gradually forgets the constraints set in the first half during the latter half of a long generation, leading to issues such as inconsistent code style and variable naming conflicts.

The project initially kept the entire plan in the conversation context, which led to drift phenomena such as a second rabbit appearing where it shouldn't, prohibited props reappearing, and scene order being rewritten on its own.