context-manager

Context Manager Skill

Manage context memories created by the PreCompact hook.

When to Activate

Use this skill when user:

• Asks about previous work ("what did we work on?", "show my contexts")
• Wants to list saved memories ("list contexts", "context history")
• Needs to load specific context ("load the auth context", "restore previous work")
• Wants to search memories ("find context about database")
• Asks about context management ("how many memories", "delete old contexts")

Directory Structure

Context memories are stored in the project's .claude/memories/ directory:

.claude/memories/
├── index.json                      # Global index of all memories
└── {context_id}/
    ├── {timestamp}/
    │   ├── memory.json            # Memory stored as JSON
    │   └── metadata.json           # Machine-readable metadata
    └── latest -> {timestamp}       # Symlink to most recent

Available Actions

1. List Contexts

Read .claude/memories/index.json and present available contexts.

Output format:

## Available Context Summaries

| # | Context ID | Date | Trigger | Files Modified |
|---|------------|------|---------|----------------|
| 1 | abc123... | 2025-11-23 19:04 | auto | 15 |
| 2 | def456... | 2025-11-22 14:30 | manual | 3 |

Total: 2 contexts stored

2. Load Context

Load a specific context memory and optionally inject into conversation.

Steps:

1. Read memory from .claude/memories/{id}/{timestamp}/memory.md
2. Read metadata from .claude/memories/{id}/{timestamp}/metadata.json
3. Present to user
4. Ask if they want it injected into current conversation

Context injection format:

<previous-context>
[Summary content here]
</previous-context>

3. Search Contexts

Search through memories by keyword or topic.

Steps:

1. Read index.json for context list
2. For each context, read memory.md
3. Search for matching keywords
4. Return ranked results

4. Context Statistics

Provide overview of stored contexts.

Output:

## Context Summary Statistics

- Total contexts: 15
- Total compactions: 23
- Most active project: /path/to/project
- Most common topics: #api, #authentication, #bugfix
- Storage used: ~2.3 MB

Tool Usage

Use these tools to implement actions:

• Read - Read index.json and memory files
• Glob - Find memory files: .claude/memories/**/*.md
• Grep - Search within memories for keywords

Response Guidelines

1. Be concise - Show tables/lists, not walls of text
2. Show recent first - Most recent contexts at top
3. Truncate IDs - Show first 8 chars of context IDs
4. Include dates - Always show human-readable dates
5. Offer actions - After listing, offer to load specific context

Error Handling

• No memories directory: "No context memories found. Summaries are created automatically when context is compacted."
• No index.json: "Summary index not found. Run /compact to create your first memory."
• Context not found: "Context '{id}' not found. Available contexts: [list]"

Integration with PreCompact Hook

This skill reads data created by the PreCompact hook, which is automatically registered when the plugin is installed. No manual configuration required.

Index.json structure:

{
  "memories": [
    {
      "context_id": "abc123...",
      "timestamp": "20251123_190448",
      "created_at": "2025-11-23T19:04:48Z",
      "trigger": "auto",
      "project": "/path/to/project",
      "files_modified": ["file1.py", "file2.ts"],
      "message_count": 150,
      "memory_path": "abc123.../20251123_190448/memory.md"
    }
  ],
  "last_context": "abc123..."
}

Example Interactions

User: "What contexts do I have?" Response: [List contexts table]

User: "Load the most recent context" Response: [Show memory, ask about context injection]

User: "Find contexts about authentication" Response: [Search and show matching contexts]

User: "How many memories are stored?" Response: [Show statistics]