# Context Anchoring Context anchoring is the practice of externalizing decision context into living documents that persist across AI-assisted development sessions. The concept was described by [[Martin Fowler]]'s team at Thoughtworks in their work on reducing friction with AI. The core problem: AI conversations are ephemeral. Decisions made early in a session fade as the [[Context Window]] fills. Research (Liu et al., 2023) shows models perform worse on information buried in the middle of conversations. Worse, the *reasoning* behind decisions fades faster than the decisions themselves. Developers end up keeping conversations running far too long, not because long sessions are productive, but because closing the session means losing context that exists nowhere else. Context anchoring solves this by capturing context in two layers: - **Priming documents** (project-level): tech stack, architecture patterns, naming conventions. Relatively stable, updated quarterly. Shared across all features. This is essentially what CLAUDE.md files and [[AI Agent Skills]] provide in [[Claude Code]] - **Feature documents** (feature-level): specific decisions, rejected alternatives, constraints, open questions, implementation state. These evolve rapidly with each session and enable coordination across developers and AI sessions The litmus test: "Could I close this conversation right now and start a new one without anxiety?" If discomfort emerges, context is trapped in the wrong medium. Context anchoring is a [[Context Hygiene]] practice that directly combats [[Context Drift]] and [[AI Context Rot]]. By externalizing context into version-controlled documents, it becomes reviewable, shareable, and durable. It's also a form of [[Progressive Disclosure]] at the project level: priming documents provide always-on context while feature documents load on demand. ## References - https://martinfowler.com/articles/reduce-friction-ai/context-anchoring.html - Liu et al. (2023). "Lost in the Middle: How Language Models Use Long Contexts" ## Related - [[Context Engineering]] - [[Context Hygiene]] - [[Context Drift]] - [[AI Context Rot]] - [[Context Window]] - [[Claude Code Memory]] - [[AI Agent Skills]] - [[Progressive Disclosure]] - [[Prompt Lazy Loading AI Design Pattern (PLL)]] - [[Martin Fowler]]