• GIF

  Summarizing Current Page (Automatically)

• GIF

  AI-Parsed Input Elements Realtime Popup

• 

• 

• 

• 

• 

• 

Semantic Memory Extension — Project Story

About the Project

The Semantic Memory Extension was created for the Google Chrome Built-In AI Hackathon as an exploration of what’s possible when on-device AI becomes a first-class citizen in the browser. The idea was simple but ambitious:

“What if your browser could remember semantically relevant notes — locally — and bring them up just when you need them?”

From that idea came a privacy-first Chrome MV3 extension that stores “memories” (semantic text snippets) in IndexedDB, retrieves them using on-device embeddings (via the Prompt API / Gemini Nano), and optionally integrates with OpenRouter for advanced cloud models. It features a popup workflow, a Memory Dashboard, and a floating AI chat for natural-language memory management — all unpacked, all open.

Inspiration

Modern browsers are full of forms, prompts, and repetitive input contexts — but our notes live elsewhere: Notion, Google Docs, or random scratchpads. I wanted a tool that:

• Understood what the page was asking for, semantically;
• Surfaced contextually relevant knowledge (like code snippets, country lists, or prior messages);
• And did so without sending anything to a remote server.

The hackathon’s emphasis on Chrome’s Built-In AI (Gemini Nano) provided the perfect opportunity: everything could stay local, private, and fast.

How It Was Built

Architecture

The extension consists of four main parts:

1. Content Script

   • Extracts visible form fields (<input>, <textarea>, <select>).
   • Uses label hierarchy and nearby DOM text to infer intent.

2. Background Service Worker

   • Handles the memory store (IndexedDB).
   • Runs embedding and similarity search (cosine distance; optional WebGPU acceleration).
   • Routes requests to providers (Built-In AI, OpenRouter, or Local Minimal).

3. Popup UI

   • Popup-first experience: when opened, it analyzes the page, builds a query, retrieves top-K memories, and shows them instantly.
   • Skeleton loaders provide feedback while data streams in.

4. Dashboard (Options Page)

   • Memory management (CRUD, import/export).
   • PDF/DOCX parsing via PDF.js and js-word.
   • Floating AI Chat that manipulates memories through natural language commands. ### Privacy Design

Default mode: Chrome’s on-device AI (Gemini Nano).

Local Privacy Mode: hard-disable all network calls; embeddings via Local Minimal.

Cloud Option: OpenRouter API (user-supplied key) with model picker.

MV3 constraints: no remote code, static UI, strict CSP.

Outcome

• The result is a prototype Chrome extension that:

• Opens instantly via popup;

• Understands your current page;

• Suggests relevant memories semantically;

• Lets you chat with your memory base via on-device or cloud AI;

• And remains 100 % client-side and privacy-respecting.

Built With

• promptapi