# Context Inertia Context inertia is when an AI agent hyper-focuses on a hypothesis, a subtask or a goal, and keeps pushing on it long after it should have stopped. The longer it works on something, the more its context is full of that something, and the harder it becomes to see anything else. Drew Breunig named it in 2026 as one of two new failure modes of long-horizon agents (the other is [[Fighting the Weights]]). It's the long-run cousin of the classic failures ([[Context Poisoning]], [[Context Distraction]], [[Context Confusion]], [[Context Clash]]): nothing in the context has to be wrong. The problem is momentum. ## What it looks like - **Anthropic's Memento experiments** (AI replication research): every ~90 minutes the agent writes itself a short note, then its context is cleared while the environment stays as it was. In one run, a Qwen agent spent hours fixing package versions when the real bug was an endpoint that didn't match the grader's API - **Fork-and-Flush** (Cai and Ziakas, Microsoft Research): autoresearch runs plateau after early gains. Agents make smaller and smaller changes to the same idea. Final scores ranged from 68.3 to 100 depending on how stuck the run got - **"Scaling Discovery Requires Forgetting"** (Shangyin Tan, UC Berkeley): on the AHC014 optimization task, a team of 4 agents that periodically cleared their context (but kept their artifacts) scored 1,144,695. Four parallel agents without forgetting got 1,007,073; a single agent 961,136 The Gemini 2.5 technical report saw an early version of this: past roughly 100K tokens, the Pokémon-playing agent favored repeating actions from its history over making new plans. ## How to fight it - **Reset the context, keep the artifacts.** Code, notes, design docs and plans survive; the transcript doesn't. A fresh agent reading the artifacts sees the situation without the history of failed attempts. Breunig starts to worry around 200K tokens and resets before big jobs (others report no degradation until 700-800K with recent models, so treat the number as a habit, not a law) - **Write a short handoff note before clearing**, like Memento. The note forces a summary of what's known, which is often where the agent notices it's stuck - **[[Context Isolation]] through subagents**: a fresh subagent with a clean history explores a new hypothesis without the main agent's momentum - **Offload state to files** so a reset costs little. This is why keeping a task list in a file "so progress survives compaction" is good advice beyond compaction ## Why it matters Context inertia flips an old intuition. We used to think more context was the goal, and forgetting the problem. For long-running agents, forgetting on purpose is part of the method. See [[Strategic Forgetting]]. It's also a human failure. Anyone who has debugged the wrong thing for three hours, then found the bug in ten minutes after a night's sleep, has met context inertia without a model involved. ## References - Drew Breunig, "The State of Context Engineering" (2026-10-24): https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html - Hacker News discussion: https://news.ycombinator.com/item?id=50235780 ## Related - [[Context Engineering]] - [[Strategic Forgetting]] - [[Fighting the Weights]] - [[Context Isolation]] - [[AI Subagents]] - [[AI Context Rot]] - [[AI Instruction Drift]] - [[Context Poisoning]] - [[Loop Engineering]]