## Definition **Context rot** is the degradation of a model's accuracy and recall as its [[Context Window]] fills. The blunt lesson: more context is not automatically better. Past some point, adding material *lowers* answer quality rather than raising it. ## The counterintuitive shape Naive intuition says a bigger window is strictly better — more facts, more recall. Reality is non-monotonic. As the window approaches capacity, the model attends less reliably to any given fact, conflates similar passages, and loses the thread of early instructions. A 200K-token window stuffed to 190K can answer *worse* than the same task framed in 20K of curated context. ``` quality ^ .--''''--. | .-' '-. | / '-.____ <- rot | / +------------------------------> tokens in window ``` ## Broader than lost-in-the-middle The [[Lost in the Middle Effect]] is a *positional* phenomenon: facts in the middle of a long input are recalled worse than facts at the ends. Context rot is the larger family. Lost-in-the-middle is one mechanism by which rot manifests, but rot also includes interference between similar items, dilution of instructions, and the model spending attention on irrelevant bulk. Position is one axis; volume and noise are others. | | Lost in the middle | Context rot | |---|---|---| | Cause | Position in window | Window fill + noise | | Scope | Special case | General degradation | | Fix | Reorder key facts | Curate, don't cram | ## The rationale for curation Context rot is *the* argument for treating the window as a scarce, curated resource rather than a dumping ground. Every token you add competes for attention with every other token. This motivates the whole discipline downstream: [[Context Compaction]] to summarize drift, [[Context Editing]] to drop spent payloads, and a [[Layered Memory Architecture]] to keep durable facts *out* of the window until needed. ## Practical implications - Retrieve the *relevant* documents, not all plausibly relevant ones. - Drop tool outputs once acted upon. - Put the spec and the current goal where the model attends best — typically near the boundaries, not buried mid-context. - Measure quality against window fill; do not assume the big window is free. ## Related - [[Context Window]] - [[Lost in the Middle Effect]] - [[Context Compaction]] - [[Context Editing]] - [[Layered Memory Architecture]]