MaxMi
A macOS menu bar app that passively builds an ambient memory of what you read, stores it locally with on-device vector search, and exposes it to Claude over MCP.
A native macOS ambient-memory tool. It watches the focused browser window through the Accessibility APIs, snapshots page content, dedupes and versions it, then runs Gemini fact extraction and per-fact embeddings into an on-device vector index. Nothing leaves the machine except the extraction calls.
The storage layer is GRDB over SQLite with a vendored sqlite-vec C extension compiled in for local semantic search, so retrieval needs no server. Memory content is encrypted at rest with AES-256-GCM (an enc:v1: wire format) and the key lives in the login Keychain; metadata, URLs, and embeddings stay cleartext on purpose so search and dedup stay fast.
It's structured as a layered SwiftPM package: a Core target for the capture pipeline and crypto, a Capture target for the AX reader and browser extraction, a Store target for the GRDB database and vector index, and a Relay target for the Gemini client with a streaming JSON parser. A separate maxmi-mcp executable speaks JSON-RPC and opens the same database read-only so it never interferes with capture, exposing semantic search_memory and thread tools to any MCP client like Claude.