Claude Code's memory docs push hard rules out of CLAUDE.md and into a PreToolUse hook

Both files arrive as prompt text. The documentation is explicit about what that buys: "Claude treats them as context, not enforced configuration." ^[3] A rule written in CLAUDE.md is a strong suggestion, evaluated against everything else in the window, including the user's last message.

That framing sets the size guidance. Anthropic targets under 200 lines per CLAUDE.md file and says longer files consume more context and reduce adherence ^[6]. The documentation states that effect without publishing the measurement behind it ^[20]. For 200 to be the right ceiling in your repo, your instructions would have to be about as compressible as the ones behind the guidance, and your sessions about as long. A file of 200 concrete path rules is a different object from 200 lines of style prose. On the quality side the docs are more useful than the count: "The more specific and concise your instructions, the more consistently Claude follows them." ^[5]

Imports do not fix size. A CLAUDE.md can pull in other files with `@path/to/import`, and imported files are expanded and loaded into context at launch alongside the file that references them, recursively, to a maximum depth of four hops ^[8]. Four hops is more depth than an instruction file needs. Split a 600-line instruction set into six 100-line files and launch context still holds 600 lines ^[19]. Relative paths resolve against the file containing the import, not the working directory ^[10].

Rules under `.claude/rules/` load only when Claude works with matching files ^[7]. Path scoping is the part that actually defers tokens, and a migration rule can sit out of a frontend session entirely.

A PreToolUse hook runs outside the model's judgement. The docs name it as the way to block an action regardless of what Claude decides ^[4]. In my view, any rule whose violation you would write up as an incident belongs there, and its restatement in CLAUDE.md is then documentation for the humans.

Conflict resolution is manual: the guidance is to review `.claude/rules/` periodically and remove outdated or conflicting instructions ^[15]. In a monorepo, `claudeMdExcludes` skips CLAUDE.md files from other teams that are not relevant to your work ^[16]. Auto memory adds a second writer to the same surface, since those notes come from your corrections and preferences rather than a commit you reviewed ^[14].

To confirm a file loaded at all, run `/context` in a session and check the list under Memory files ^[13].