Pensieve: Open-Source Local Indexing Tool Reduces AI Agent Tool Calls by 43%

Pensieve is a free, open-source local indexing tool designed to enhance AI coding agents by providing precise context without excessive token usage. By connecting via the Model Context Protocol (MCP), it allows agents like Claude Code, Codex, and Cursor to perform semantic and exact searches across files, repositories, and past agent sessions locally on the user's machine.

Benchmark Results

In tests using Claude Sonnet within Claude Code across four real-world repositories (ranging from 504 to 34,847 files), agents using Pensieve showed significant efficiency gains compared to standard grep-based workflows:

• Tool Call Reduction: Average tool calls per lookup dropped by 43%, from 3.4 to 2.0.
• Cost Efficiency: Lookups were 18% cheaper on average ($0.041 vs $0.050), with savings ranging from 6% to 28% depending on the repository size.
• Speed: Lookups were 21% faster.
• Accuracy: The correct file was found in 131 of 144 runs, compared to 123 with grep alone.

Key Features

• Local Search: Indexing and search run entirely on the user's Mac, ensuring privacy. Secrets and password exports are automatically skipped.
• Agent Integration: The pensieve integrate command configures agents to prioritize Pensieve searches over grep, adding specific instructions to agent configuration files.
• Knowledge Management: The tool analyzes git blame and commit history to identify code owners and flag potential knowledge gaps where active contributors are no longer present.
• Availability: The full application requires macOS 14+ on Apple Silicon. Backend components and a browser visualizer are available for Linux and Intel Macs.