What Claude Code loads

Claude Code Memory Diagnoser scans the same places Claude Code reads memory from. It shows each file in the tree with a load type, so you can see which files cost context in every session and which load only when Claude needs them.

Discovery order

The app builds the list in this order:

Managed policy CLAUDE.md
User CLAUDE.md
User rules
Project CLAUDE.md files, in the project directory and each parent directory
Nested CLAUDE.md files in subdirectories of the project
Project rules
Project skills
Auto memory
Agent memory
Files that the files above import

The tree groups the results as Managed Policy, User, Project, Rules, Auto Memory, Skills, and Agent Memory. Each group header shows the number of top-level files in the group.

Load types

Each row in the tree has an icon for its load type. Hover the icon to see its name.

Icon	Load type	When Claude Code loads the file
●	Always	In every session.
◒	Startup	At session start, up to a size cap.
○	Conditional	When Claude works on a file that matches the rule’s `paths` globs.
◌	On-demand	Only when Claude reads it, for example a skill body or a memory topic file.
◎	Tree	When Claude works in that subdirectory.
@	Import	With the file that imports it.
↗	Link	Not loaded automatically. A markdown link only points to the file.

The tree shows conditional, on-demand, tree, and link rows in a muted color.

Config directory

User-level files live in the Claude config directory, ~/.claude by default. This page writes <config dir> for that directory. To use a different one, see Config dir.

Managed policy

An administrator can put a CLAUDE.md at a fixed system path. Its load type is always.

OS	Path
Windows	`%ProgramFiles%\ClaudeCode\CLAUDE.md`
macOS	`/Library/Application Support/ClaudeCode/CLAUDE.md`
Linux	`/etc/claude-code/CLAUDE.md`

User CLAUDE.md

Your personal instructions for all projects live in <config dir>/CLAUDE.md. Its load type is always.

Project CLAUDE.md

The app starts at the project directory and goes up one directory at a time. It stops before the filesystem root. In each directory it checks four names:

CLAUDE.md
.claude/CLAUDE.md
CLAUDE.local.md
.claude/CLAUDE.local.md

All of these have the load type always. CLAUDE.local.md files show in the Project group. There is no separate group for local files.

Nested CLAUDE.md

The app also looks for the same four names in subdirectories of the project, up to 6 levels deep. The tree shows each one with its relative path, for example src/api/CLAUDE.md. Their load type is tree. Claude Code loads them when it works in that directory.

The scan skips node_modules, .git, .hg, .svn, dist, build, out, .next, .nuxt, vendor, __pycache__, .venv, venv, and every directory whose name starts with a dot.

Rules

Rules are .md files in two folders. The app reads subfolders too.

<config dir>/rules/ for user rules
<project>/.claude/rules/ for project rules

A rule with a non-empty paths list in its frontmatter is conditional. A rule without paths is always loaded. The app does not scan .claude/rules in parent directories of the project.

Project skills

The app lists each SKILL.md under <project>/.claude/skills/, up to 4 levels deep. It also finds .claude/skills folders in project subdirectories and names those skills <subdir>:<name>. The skill name comes from the name frontmatter field, or else from the folder name.

Skills are on-demand. Claude Code loads a skill body only when it uses the skill. The app does not scan user skills in <config dir>/skills.

Auto memory

Claude Code keeps auto memory for each project in <projects base>/<encoded path>/memory/.

The projects base is <config dir>/projects. If autoMemoryDirectory is set in <config dir>/settings.json, the app uses that folder instead. It reads this setting from user settings only, not from project settings.

To encode the project path, the app changes each \, /, and : to -. For example, C:\Users\foo becomes C--Users-foo.

If that folder has no memory directory, the app tries two more lookups:

If the project is a git linked worktree, it uses the memory of the main worktree.
It uses the first folder in the projects base whose name ends with the last two segments of the project path. This match ignores case.

In the memory folder, MEMORY.md has the load type startup. Claude Code reads only its first 200 lines, up to 25 KB. See Read a MEMORY.md index for how the file view shows this. The other .md files in the folder are on-demand topic files.

Agent memory

Subagents with persistent memory keep it in one folder per agent:

Scope	Folder
User	`<config dir>/agent-memory/<agent>/`
Project	`<project>/.claude/agent-memory/<agent>/`
Local	`<project>/.claude/agent-memory-local/<agent>/`

Each folder has the same layout as auto memory: MEMORY.md loads at startup and the other files are on-demand. In the tree, the Agent Memory group is collapsed at first, and each agent has a header with its scope.

Imports and links

The app reads each file for references to other files:

@path/to/file.md is an import. The path must contain a / and the last segment must have a file extension, so @biomejs/biome does not count.
@~/path/file.md is an import from your home directory.
A standalone @name.md is an import of a file next to the current one.
[text](file.md) is a link. Links to http and https URLs do not count.

A path that starts with ~ resolves from your home directory. Other paths resolve from the folder of the file that holds the reference.

Imported files show under their parent with the load type import. Linked files show under their parent with the load type link. Topic files in a memory folder that MEMORY.md links to show under MEMORY.md and keep the on-demand type. The app follows imports and links up to 5 levels deep. When a target file does not exist, the file view lists the reference as unresolved and marks it with ⚠.

Footprint

The Memory sources bar at the top and the summary cards at the bottom show the standing cost of memory. That is the set of files that can load in a session without Claude reading them on purpose. The count includes conditional rules, which load only when Claude works on a file that matches their paths. The count leaves out:

Skill bodies
Files reached only through a markdown link
Nested CLAUDE.md files with the tree load type, and the files they import
On-demand topic files in auto memory and agent memory

For each skill, the count adds the characters of its name and description, because those go into every session. A skill with disable-model-invocation: true adds nothing.

The bar shows characters and bytes, with one colored segment per group. The cards show Files, Chars, Bytes, and Always loaded. Always loaded counts files with the always or startup load type. The app does not estimate tokens. To use the bar, see See the memory footprint.

Refresh

The app keeps the scan result for 30 seconds. It scans again when you switch projects, delete a file, or clean up orphaned links. To see a change you made in an editor at once, press r or click Refresh in the top bar.

To scan a different project, click the folder button in the top bar or press Shift+P. See Switch projects.