Inspiration

Every AI agent starts from zero. A user's preferences, project constraints, writing style, travel habits, and private notes get trapped inside separate tools like ChatGPT, Claude, Gemini, Codex, and future agents. Memory Dock gives the user a central memory layer they own.

What it does

Memory Dock is a native macOS menu bar app that works like a Dropbox for agent memory. Users drop files or paste notes into the menu bar app. Gemini extracts durable memories and maintains a structured LLM Wiki. Elastic indexes raw sources, extracted facts, wiki pages, context packs, and audit events.

The user can search their memory and export scoped context packs for Gemini, Claude, ChatGPT, Codex, or JSON. Sensitive and never-export memories are excluded from portable packs by default.

How we built it

The project has a SwiftUI MenuBarExtra macOS app, a Node.js/TypeScript API, Gemini-powered memory extraction, an Elastic-backed memory store, deterministic demo data, and Cloud Run deployment docs.

The app supports text, Markdown, JSON, and simple PDF ingestion. The API creates memory sources, atomic facts, LLM Wiki pages, context packs, and audit records. The same code runs in local demo mode or Elastic Cloud mode.

How we used Elastic

Elastic is the load-bearing retrieval and governance layer. Memory Dock stores sources, facts, wiki pages, context packs, and audit logs in Elastic indices. The Gemini proof agent retrieves scoped memories through Elastic MCP, cites memory IDs, and avoids exporting sensitive scopes.

How we used Google Cloud and Gemini

Gemini turns dropped files and notes into durable memories and LLM Wiki updates. Google Cloud Run hosts the API. The planned Memory Passport Agent in Gemini/Agent Builder retrieves memory through Elastic MCP and creates privacy-safe context packs for other agents.

Challenges

The main challenge was turning the idea from a normal RAG app into a compounding memory system. Instead of only retrieving raw chunks, Memory Dock maintains durable facts, wiki pages, source links, privacy scopes, and audit trails.

What is next

Next steps are direct chatbot integrations, richer PDF and screenshot ingestion, contradiction detection, Obsidian export, and per-agent memory permissions.

Built With

Share this project:

Updates