Context explained: what Claude knows right now

Anthropic's documentation defines the context window as all the text a model can reference when generating a response, including the response itself, and distinguishes it from the large corpus the model was trained on. It describes it as a working memory. Everything in the request counts: the system prompt, every message (including tool results, images and documents), tool definitions, and the output including extended thinking.

The window is up to 1M tokens on the current Claude models. Claude Code's docs add that Fable models, Sonnet 5 and later, Haiku 5.5, Opus 4.6 and later and Sonnet 4.6 support a 1M window, with availability depending on plan.

Each of these costs context on every request. Anthropic's extension guide gives the loading rules: CLAUDE.md loads in full at session start; skills load only their descriptions at start and the full body when used; MCP servers load tool names with schemas deferred until needed (tool search is on by default); hooks cost nothing unless they return output; subagents get their own isolated context. You can see the real numbers with the /context command, which gives a live breakdown by category, and open memory files with /memory.

Anthropic's documentation says more context is not automatically better: as token count grows, accuracy and recall degrade, a phenomenon it calls context rot, which makes curating what is in context as important as how much space is available. It also notes that chat interfaces such as claude.ai can manage the window on a rolling first-in, first-out basis, so very old material may simply fall off the desk.

Per Anthropic's context-window page, after compaction: the system prompt and output style still apply; project-root CLAUDE.md and unscoped rules are re-injected from disk; auto memory is re-injected from disk; skill descriptions are rebuilt and re-injected (requires Claude Code v2.1.304 or later); Claude re-reads up to five of the most recently modified files, and a file over 5,000 tokens comes back as a path reference without content; invoked skill bodies are re-injected but capped at 5,000 tokens per skill and 25,000 total, oldest dropped first (so put the most important instructions near the top of SKILL.md); rules with paths: frontmatter and nested CLAUDE.md files reload on demand; background commands keep running.

The memory page adds the practical lesson: if an instruction disappeared after compaction, it was given only in conversation, so add it to CLAUDE.md to make it persist.

/autocompact with a token count sets how full the window gets before the automatic pass runs. The documentation says to compact with a focus before a long new task so the summary keeps what you choose, and to clear between tasks because old conversation "crowds out" the files you need next and costs tokens on every message. For API users, server-side compaction (beta) and context editing (clearing old tool results or thinking blocks) serve the same purpose, and the token counting endpoint estimates a request before you send it.