mem-search

Memory Search

Search is automatic. The UserPromptSubmit hook detects history-related prompts and injects relevant context.

Just Ask Naturally

The hook recognizes patterns like:

• "What did we do last session?"
• "Did we fix this bug before?"
• "How did we implement X last time?"
• "What changes were made to this file?"
• "Why did we choose Postgres?" (finds preloaded decisions)

When detected, it performs semantic search (if model installed) or text search and injects matching events as context.

How It Works

1. SessionStart - Recent events injected, preload files indexed
2. UserPromptSubmit - History queries trigger semantic search (includes preloaded knowledge)
3. PostToolUse - All tool executions recorded for future sessions

Preloaded Knowledge

Projects can include searchable documentation in .claude-mem/preload/:

project/
├── .claude-mem/
│   └── preload/
│       ├── architecture.md     # System design
│       ├── conventions.md      # Code conventions
│       └── decisions/          # Subdirs = categories
│           └── database.md     # Why we chose X

This content is:

• Indexed on session start (not auto-injected into context)
• Found via semantic search when you ask relevant questions
• Re-imported when files change (hash-based detection)

Setup Semantic Search

For better search results, install the embedding model:

npm run setup:model  # Downloads ~24MB model

Without it, search falls back to LIKE-based text matching.

Manual Browsing

To browse history manually, start the viewer:

npm run viewer  # Opens http://localhost:37777

The viewer provides:

• /api/observations - Recent tool events
• /api/search?q=query - Search events
• /api/projects - List of projects
• /api/stats - Database statistics

Privacy

Wrap sensitive content with <private> tags to exclude from storage:

<private>my secret API key</private>