Minimi

Product Introduction

1. Definition: Minimi is a native Mac desktop application that functions as an ambient, on-device context engine for Claude. It is a background utility that captures a continuous stream of user activity across the system—including web tabs, documents, calls, messages, and applications—and structures this data as live, queryable memory for integration with the Claude AI model.
2. Core Value Proposition: Minimi eliminates the "context gap" in AI interaction by providing Claude with full, passive awareness of your digital workday. Its primary purpose is to deliver "zero-prompt" contextual awareness, enabling Claude to answer questions about past activities, meetings, and research without the user needing to manually compile or describe background information. This enhances productivity and ensures more accurate, personalized responses from Claude.

Main Features

1. Ambient Context Capture & Memory Indexing: The application operates silently in the Mac menu bar ("Install & Forget"), continuously capturing structured data from open applications and browser tabs. This includes text from Slack threads, documents, call transcripts, and web pages. The captured content is converted into vector embeddings using a paid, secure Gemini API endpoint and stored in a local, on-device vector database. This creates a searchable, long-term memory of your digital activity.
2. Claude MCP Integration: Minimi connects to Claude via the Model Context Protocol (MCP). The user installs the app, copies a unique MCP link, and pastes it as a custom connector in Claude's interface. Once connected, Minimi functions as a live data source that Claude can query in natural language to retrieve specific past context during a chat session.
3. On-Device Privacy Architecture: The entire memory pipeline is designed for privacy. All data processing, embedding generation, and storage occur locally on the user's Mac. The product states that the Gemini API is used only on a paid plan with no access to user data, and data in transit to and from the LLM is encrypted, being decrypted only momentarily for processing. This positions it as a privacy-first alternative to cloud-based memory solutions.

Problems Solved

1. Pain Point: Contextual Amnesia in AI Conversations. Users frequently struggle to provide Claude with the full background of a question, leading to generic or incomplete answers. Minimi solves the problem of "context window limitation" and "manual briefing fatigue" by automatically feeding relevant historical and real-time data to the AI.
2. Target Audience: Knowledge workers, software developers, project managers, writers, researchers, and power users of Claude on Mac. Specifically, professionals who engage in long-form projects, cross-functional communication, and deep research where recalling details from past days, weeks, or meetings is critical.
3. Use Cases:
   • Project Management: Instantly retrieving action items, decisions, and discussions from past Slack threads and meeting calls without searching.
   • Research & Writing: Having Claude synthesize information scattered across multiple browser tabs and documents into a cohesive summary.
   • Communication Recall: Answering questions like "Who sent me the screenshot?" or "What did I promise to send by Friday?" with precise context.
   • Productivity Analysis: Asking Claude to analyze personal work patterns, such as "What have I been quietly avoiding?" or "What should I focus on right now?" based on captured activity.

Unique Advantages

1. Differentiation: Minimi differentiates itself from standard AI assistants and basic note-taking apps by offering passive, system-wide capture specifically optimized for Claude. Unlike cloud services (e.g., note-taking plugins), it operates entirely on-device, offering superior privacy. The key differentiator is its focus on ambient intelligence—it works without user intervention, creating a memory that is always up-to-date.
2. Key Innovation: The core innovation is the privacy-preserving, local vector database architecture for continuous context. By performing all embedding and storage on the Mac and using the Gemini API only for secure, non-accessible processing, it achieves a high level of memory accuracy (cited as 54% on the BEAM benchmark) without compromising user privacy. The seamless MCP integration makes this advanced memory system directly accessible within Claude's interface.

Frequently Asked Asked Questions (FAQ)

1. How does Minimi work with Claude, and is it secure? Minimi connects to Claude using the Model Context Protocol (MCP) link provided by the app. All captured context is processed, embedded, and stored locally on your Mac. Data is only sent to Claude for processing when you initiate a query, and it is encrypted in transit. The system is designed for on-device privacy, with no cloud storage of your memory data.
2. What kind of data does Minimi capture, and can I control it? Minimi captures content from active windows including web browsers, documents, messaging apps (like Slack), and call transcripts. The product describes it as capturing "everything you have read, said, or heard." Control and permission details would be found in the app's privacy policy, but the core principle is system-wide capture for context.
3. Is Minimi better than simply using Claude's projects or manually typing context? Yes, Minimi solves a different problem. It provides automatic, continuous, and comprehensive context that is impractical to compile manually. While Claude's projects are for organizing documents you explicitly add, Minimi creates a live, searchable memory of your natural digital activity, answering questions about the past without preparation.
4. What platforms does Minimi support, and what are its system requirements? Minimi is a Mac-only application at launch. The core system requirement is a macOS device to run the on-device memory and vector database. It is designed to be a lightweight, "install & forget" utility that runs in the background.
5. How does the on-device memory accuracy compare to other solutions? According to the product, Minimi's memory retrieval achieves 54% accuracy on the BEAM (ICLR 2026) benchmark for long-term memory. This is stated to be 36% more accurate than the previous state-of-the-art (SOTA) method called LIGHT (BEAM), highlighting its technical advancement in accurate memory recall.