auto-memory

claude-code-auto-memory

Your CLAUDE.md, always in sync. Minimal tokens. Zero config. Just works.

A Claude Code plugin that watches what Claude Code edits, deletes, and moves - then quietly updates your project memory in the background. No manual maintenance needed.

The Problem

CLAUDE.md files become stale as codebases evolve:

• Build commands change but memory stays outdated
• Architecture shifts go unrecorded
• Code conventions drift without memory updates
• New team members get incorrect context

Manual maintenance is tedious and often forgotten.

The Solution

claude-code-auto-memory automatically updates CLAUDE.md when Claude Code makes changes. Processing happens in an isolated agent, so it doesn't consume your main conversation's context window.

Claude Code edits code -> Plugin tracks changes -> Isolated agent updates memory -> Context stays fresh

Features

• Automatic sync: Tracks Edit/Write/Bash operations and updates CLAUDE.md at end of turn
• Bash operation tracking: Detects rm, mv, git rm, git mv, unlink commands
• Minimal-token tracking: PostToolUse hook has no output; stop hook triggers isolated agent
• Isolated processing: Agent runs in separate context window, doesn't consume main session tokens
• Marker-based updates: Only modifies AUTO-MANAGED sections, preserves manual content
• Subtree support: Hierarchical CLAUDE.md for monorepos

Installation

From Marketplace

claude plugin marketplace add severity1/severity1-marketplace
claude plugin install auto-memory@severity1-marketplace

Local Development

# Add local marketplace
claude plugin marketplace add /path/to/claude-code-auto-memory/.dev-marketplace/.claude-plugin/marketplace.json

# Install from local
claude plugin install auto-memory@local-dev

Commands

/auto-memory:init

Initialize CLAUDE.md structure for your project with an interactive wizard.

/auto-memory:init

The wizard will:

1. Analyze your codebase structure
2. Detect frameworks and build commands
3. Identify subtree candidates (for monorepos)
4. Present findings for your approval
5. Generate CLAUDE.md with auto-managed sections

/auto-memory:calibrate

Force a full recalibration of all CLAUDE.md files.

/auto-memory:calibrate

/auto-memory:sync

Sync CLAUDE.md with manual file changes detected by git. Use this when you've edited files outside Claude Code (in your IDE, terminal, etc.) and want to update memory without a full recalibration.

/auto-memory:sync

Detects:

• Modified tracked files (git diff)
• Staged files (git diff --cached)
• New untracked files (git ls-files --others)

/auto-memory:status

Show current sync status and pending changes.

/auto-memory:status

How It Works

sequenceDiagram
    participant User
    participant Claude as Claude Code
    participant PostHook as PostToolUse Hook
    participant DirtyFiles as dirty-files
    participant StopHook as Stop Hook
    participant Agent as memory-updater Agent
    participant Skill as memory-processor Skill

    User->>Claude: Edit files
    Claude->>PostHook: Edit/Write/Bash operation
    PostHook->>DirtyFiles: Append file path (zero tokens)
    Note over PostHook: No output - pure tracking

    User->>Claude: Continue working...
    Claude->>PostHook: More edits
    PostHook->>DirtyFiles: Append paths

    Note over Claude: Turn ends
    Claude->>StopHook: End of turn
    StopHook->>DirtyFiles: Check for dirty files

    alt Dirty files exist
        StopHook->>Agent: Spawn isolated agent
        Agent->>Skill: Invoke memory-processor
        Skill->>Skill: Analyze changes
        Skill->>Skill: Update CLAUDE.md sections
        Agent-->>StopHook: Complete
        StopHook->>DirtyFiles: Clear processed files
    else No dirty files
        Note over StopHook: No action needed
    end

Architecture Overview

PostToolUse Hook (Edit|Write|Bash)
    |
    v (append file paths)
.claude/auto-memory/dirty-files
    |
    v (end of turn)
Stop Hook
    |
    v (spawn agent)
memory-updater Agent (isolated context)
    |
    v (invoke skill)
memory-processor Skill
    |
    v
CLAUDE.md updated

Token Efficiency

• PostToolUse hook: Zero token cost (no output)
• Stop hook: Minimal output only when dirty files exist
• Agent: Runs in isolated context window - doesn't consume main session tokens
• Skills: Progressive disclosure - load only when invoked

CLAUDE.md Format

Auto-managed sections use HTML comment markers: