Summary

silverstein/minutes provides a privacy-first, open-source conversation memory layer. It captures human interactions like meetings and voice notes, making them searchable and queryable for AI agents and users. The system enhances AI context and personal recall by indexing decisions, intent, and context often missed by agents.

How It Works

The core pipeline processes audio through transcription (whisper.cpp), speaker diarization (pyannote-rs), and summarization (local or cloud LLMs). It generates structured Markdown with YAML frontmatter, detailing decisions, action items, and people. A SQLite index builds a relationship graph, enabling cross-meeting intelligence and instant queries on commitments and topics. The entire process prioritizes local execution, ensuring data privacy.

Quick Start & Requirements

• Installation: macOS users can install the desktop app via brew install --cask silverstein/tap/minutes or the CLI with brew install minutes. Cross-platform source builds require Rust and cmake (cargo install minutes-cli). An MCP server is available via npx minutes-mcp.
• Prerequisites: Rust and cmake for source builds. Optional GPU acceleration (NVIDIA CUDA, Apple Metal/CoreML) for faster transcription. LLM integration (Claude, Ollama, OpenAI) is required for summarization.
• Setup: minutes setup --model tiny downloads necessary models (75MB).
• Docs: Links to release notes are provided, but a central documentation portal is not explicitly linked.

Highlighted Details

• Privacy-Centric: All processing, including transcription and diarization, runs locally by default. Audio data remains on the user's machine.
• Relationship Intelligence: Automatically tracks and indexes people, commitments, and topics across all recorded conversations, enabling queries like "What did I promise Sarah?".
• Cross-Meeting Search: Facilitates semantic and full-text search across the entire meeting history.
• AI Integration: Offers native integration with the Claude ecosystem via MCP, allowing AI assistants to query meeting data without API keys.
• Voice Memo Pipeline: Efficiently processes short voice memos (<2 minutes) with a streamlined pipeline.
• Structured Output: Generates Markdown files with detailed YAML frontmatter for meetings, decisions, and action items, compatible with tools like Obsidian and Logseq.

Maintenance & Community

The project is maintained by Mat Silverstein. No specific community channels (Discord, Slack) or sponsorship details are provided in the README.

Licensing & Compatibility

• License: MIT.
• Compatibility: The permissive MIT license allows for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

Certain advanced features, such as calendar attendee integration, screen context capture, and native dictation hotkeys, are currently macOS-only. The Windows desktop installer is unsigned, requiring user discretion. Building from source necessitates specific development tools (Rust, cmake), and GPU acceleration requires compatible hardware and complex setup.