# Context Entropy Context entropy is the natural tendency of AI context to degrade toward disorder over time. Left unmanaged, any sufficiently complex context system accumulates contradictions, redundancies, outdated references, and noise until the signal-to-noise ratio drops below usefulness. This is distinct from [[AI Context Rot]] (staleness of specific entries) and [[Context Confusion]] (contradictory information). Context entropy is the system-level tendency; the second law of thermodynamics applied to AI context. Individual entries may each be fine, but the aggregate system becomes increasingly disordered. ## How entropy accumulates - **Additive bias**: it's always easier to add context than to remove it. New rules, memories, and instructions pile up without corresponding pruning - **Temporal layering**: context from different time periods coexists without clear precedence, creating implicit contradictions - **Multi-author drift**: in team settings, different people add context with different assumptions, styles, and conventions - **Tool output accumulation**: agent conversations grow; tool results, memories, and state accumulate across interactions - **Scope creep**: context expands to cover edge cases, eventually overwhelming the common cases it was designed for ## Entropy vs rot | | Context Entropy | [[AI Context Rot]] | |---|---|---| | **Scale** | System-level disorder | Individual entry staleness | | **Cause** | Accumulation without curation | Time passing without updates | | **Detection** | Hard; the system "feels" slower | Easier; specific entries are wrong | | **Fix** | Restructure, prune, reorganize | Update or remove specific entries | ## Fighting entropy Entropy cannot be eliminated, only managed. It requires ongoing energy (effort) to maintain order: - **Regular pruning**: the [[Context Lifecycle]] review phase exists specifically to combat entropy - **[[Context Budget]]**: hard limits force prioritization and prevent unbounded growth - **[[Context Layering]]**: separate concerns into layers so entropy in one layer doesn't contaminate others - **Compression and summarization**: periodically consolidate verbose context into concise equivalents - **[[Context Signal-to-Noise Ratio]]**: measure and track the ratio to detect entropy before it becomes critical ## References - ## Related - [[AI Context Rot]] - [[Context Lifecycle]] - [[Context Engineering]] - [[Context Hygiene]] - [[Context Bloat]] - [[Context Confusion]] - [[Context Budget]] - [[Context Signal-to-Noise Ratio]] - [[Context Layering]] - [[AI context is finite with diminishing returns]] - [[Natural tension between compression and context]] - [[Personal Context Management (PCM)]] - [[Harness Engineering]]