# Context Distraction Context distraction is when a context grows so long that the model over-focuses on it and neglects what it learned during training. The model stops reasoning from its knowledge and starts pattern-matching on its own history. The term comes from Drew Breunig's four long-context failure modes (June 2025), next to [[Context Poisoning]], [[Context Confusion]] and [[Context Clash]]. Distraction is about length. The content can be correct and relevant; there's just too much of it, and the history wins over the model's own judgment. (When irrelevant content steers the answer, that's [[Context Confusion]].) ## Evidence - **Gemini 2.5 playing Pokémon**: once the agent's context grew well beyond 100K tokens, it favored repeating actions from its history over making new plans. Gemini 2.5 Pro supports more than 1M tokens, so the problem showed up long before the window was full - **Databricks long-context study**: correctness started to fall around 32K tokens for Llama 3.1 405B, and earlier for smaller models - **"Do as I Say, Not as I Do"** (Camassa and Shiller, cited by Breunig in 2026): models told to answer in Russian but shown contradictory French examples followed the examples. Smaller models such as Gemma 3 12B dropped below 50% instruction-following after a few examples. Bigger models held out longer but weren't immune The practical lesson: the advertised context window is not the usable one. The distraction ceiling is much lower than the size limit. ## How it happens - **Long agent sessions**: tool calls, results and attempts pile up for hours - **Never resetting**: one conversation used for a whole day of work - **Large tool outputs kept inline**: logs, files and search results that stay in the history long after they were used - **Many-shot examples**: a long list of examples the model imitates instead of following the instruction Over many hours, distraction becomes [[Context Inertia]]: the agent locks onto what it has been doing and can't let go. ## Mitigation - **Summarize or compact** the history before it gets long ([[Context Compression]]) - **Offload** large results to files and keep only a pointer in the context - **Reset on purpose**: checkpoint the work into artifacts (code, notes, plans), then start fresh. Breunig starts to worry around 200K tokens; others see problems only much later with recent models, so test your own threshold ([[Strategic Forgetting]]) - **[[Context Isolation]]**: give long subtasks to [[AI Subagents]] so their history never reaches the main agent - **Fewer examples, better interfaces**: a well-designed tool parameter beats a long list of examples ## References - Drew Breunig, "How Long Contexts Fail" (2025-06-22): https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html - Drew Breunig, "The State of Context Engineering" (2026-10-24): https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html ## Related - [[Context Bloat]] - [[Context Engineering]] - [[AI context is finite with diminishing returns]] - [[Context Poisoning]] - [[Context Confusion]] - [[AI Context Rot]] - [[Context Hygiene]] - [[Progressive Disclosure]] - [[Prompt Lazy Loading AI Design Pattern (PLL)]] - [[Receptionist AI Design Pattern]] - [[Token Budget]] - [[Context Window]] - [[Personal Context Management (PCM)]] - [[Obsidian Starter Kit - Tutorial - Managing AI sessions]] - Operational guidance to prevent context distraction in OSK sessions - [[Context Inertia]] - [[Context Clash]] - [[Strategic Forgetting]] - [[Context Isolation]] - [[Context Compression]] - [[AI Subagents]]