Me Agent

How to Use

Read Mode (no arguments)

The top-level index (~/.claude/claude-me/corpus/ME.md) is already in your context via CLAUDE.md. Use it to identify relevant entries, then read specific topic files for full detail:

1. Check the index (already in context) for entries relevant to the current conversation
2. Read specific topic files at ~/.claude/claude-me/corpus/<category>/<file>.md
3. Apply this knowledge naturally — don't announce it, just adapt

Sync Mode (/claude-me sync)

Extract cross-project preferences from all active Claude Code memory folders:

clm sync

Report what was extracted: how many projects scanned, how many new entries added, and which categories they landed in.

Consolidate Mode (/claude-me consolidate)

Merge duplicates, resolve contradictions, prune project-specific leaks:

clm consolidate
clm consolidate "merge all PR-related entries"

Optional focus text guides Haiku to prioritize specific criteria. Report what changed.

Costs Mode (/claude-me costs)

Show accumulated Haiku API cost summary:

clm costs

Note Mode (/claude-me note "...")

Add a preference note to be processed on next sync:

clm note "always run tests before committing"
clm note "always run tests before committing" --now            # blocks until processed
clm note "always run tests before committing" --now --detach   # processes in background

Without --now, the note is just saved to disk (instant). With --now, add --detach to avoid blocking the session.

Interview Mode (/claude-me interview)

Present pending interview questions to the user and process their answers. Questions are generated during consolidation when Haiku encounters conflicts or ambiguities it cannot resolve alone.

1. Read ~/.claude/claude-me/pending-questions.json
2. If no questions, tell the user there are no pending questions
3. Present each question conversationally — show the question, context, and related entries
4. For each answer the user gives, run: clm note "Re: <question> — <answer>" --now --detach
5. After each answer, clear it: clm interview --clear <question-id>
6. After all questions are answered, or to clear all at once: clm interview --clear-all

Status Mode (/claude-me status)

Show corpus stats and system status:

clm status

Open / Verify / Delete (single-entry actions)

Direct actions on individual corpus entries — use when the user says "open my corpus", "this is still true", or "delete this rule":

clm open                                       # open corpus dir in VS Code
clm verify rules/never-commit-untested.md      # bump last_verified + verify_count
clm delete rules/old-rule.md                   # soft-delete (recoverable for 7 days)
clm delete rules/old-rule.md --yes             # skip confirmation

Soft-deleted entries land in ~/.claude/claude-me/trash/ and auto-prune after 7 days. Recoverable via mv until then.

Daemon Mode (/claude-me daemon ...)

Optional daily notification daemon. Surfaces one corpus entry per day at 9am for the user to review:

clm daemon enable    # register LaunchAgent, fire one test notification
clm daemon disable   # unregister
clm daemon test      # fire one notification right now
clm daemon status    # show registration state

Disabled by default — the user opts in via clm daemon enable. Requires terminal-notifier (brew install terminal-notifier).

Corpus Structure

The corpus lives at ~/.claude/claude-me/corpus/ (outside the skill repo — your personal data stays private):

~/.claude/claude-me/corpus/
  ME.md                     Top-level index
  interaction-style/        How you talk to Claude Code
    ME.md + topic files
  projects/                 What you're building
    ME.md + topic files
  rules/                    Rules you enforce
    ME.md + topic files
  patterns/                 Workflow habits
    ME.md + topic files

Each topic file has YAML frontmatter (name, description) and markdown content with optional **Why:** and **How to apply:** sections.