Context management | Operating Manual

The commands for managing the live context window and the files that persist between sessions. For how the layers fit together and what belongs where, see Memory & Context.

How you use it day-to-day

/compact [instructions] summarizes the conversation when the window fills, and you can steer what survives: /compact focus on the auth refactor and the files still to migrate keeps your working set instead of a generic summary. The highest-value long-session move.
/clear wipes the conversation context for a clean start, while leaving your CLAUDE.md and memory files in place. Reach for it between unrelated tasks rather than letting an old thread bleed into a new one.
/memory opens and edits your CLAUDE.md memory files and controls auto-memory from inside the session, so you never hand-edit the files.
# mid-session writes a one-line rule straight into CLAUDE.md instead of forgetting it after the turn.
/context in a fresh session shows what's already loaded (CLAUDE.md, MCP tools, skills). Audit it once per project and disable unused MCP servers with /mcp; startup context is a tax on every single turn.
Set /model and /effort once, at the start. Changing either mid-conversation breaks the prompt cache and re-prefills the whole window at full price. Anthropic's cost ranking: model choice first, session length second, cache breaks third, command noise fourth.
/rewind beats /compact for a wrong turn. Rewind only cuts the end and keeps the cache; save /compact for when the whole conversation needs rewriting.

Gotchas

Auto-memory loads in full every session. The MEMORY.md index is read every time, so keep it lean and one fact per entry. Past a couple hundred lines it crowds out real work.
The automatic compaction summary is generic. On a long session, run a directed /compact before the window forces one on you, or you lose the details you cared about. Compact before a break, not after: the prompt cache expires after about an hour on subscriptions, so summarizing while it's still warm is materially cheaper than resuming cold.
Noisy commands quietly eat the window. Add quiet flags (--reporter=dot and friends) or run loud commands in a subagent so pages of test output don't ride along on every turn.
A subagent resets the clock. Quality degrades as the window fills; delegating a noisy task to a subagent keeps your main context clean (it gets a fresh window).
Nothing persists unless it is in a file. Between sessions, only CLAUDE.md and memory files survive. If you want it next time, write it down (or press #).