Guides for Claude Code: memory, sessions, subagents, permissions and hooks
Claude Code rarely fails loudly. A rule does not load, a session is not where you left it, a subagent starts without the context you assumed. Each guide begins with a symptom you can see in a session and works down to the layer that actually decides the outcome, which is often not the file you edited.
Checked against each tool's documentation on 26 October 2026.
Memory and instructions#
- Claude Code auto memory — where Claude keeps its own notes, what the index loads, and how to check that it reached the session.
- Why your nested CLAUDE.md or path-scoped rule never loaded — the lazy-load trigger, why Auto mode can skip it, and how to prove what loaded.
- Claude Code loaded your CLAUDE.md but isn't following it — loaded is not obeyed: why long, conflicting or vague rules lose, and what to use when a rule must hold.
Sessions#
- Find and resume the Claude Code session you actually want — how
--continueand--resumediffer, what the picker shows, and where transcripts live.
Subagents#
- What a Claude Code subagent can and can't see — fresh context, missing history, and the MCP tool-definition budget that can still exhaust a custom subagent's window before its first turn.
Permissions and hooks#
- Why Claude Code keeps asking for permission — how allow rules match compound commands, and which layer to fix.
- Claude Code hooks that don't fire — a debugging order for hooks that run but never reach Claude, or never run at all.
The wider picture#
Persistent memory for Claude Code and Codex is the guide to instruction files, auto memory, MCP memory servers and shared stores. Context compaction covers what a long session drops, and what survives. Instruction files, skills and memory compares what each agent reads. The same problems in other agents: Codex guides, OpenCode guides, GitHub Copilot CLI guides, Gemini CLI guides, Qwen Code guides and Cursor guides.