Browser Memory Extension

Inspiration

In today's AI landscape, a fundamental problem persists: every AI conversation operates in isolation. When you switch between tools or start a new chat, you lose all context—forcing you to repeatedly explain your work, projects, and preferences. This isn't just inconvenient; it represents a massive productivity drain as knowledge workers waste hours daily re-explaining themselves to different AI systems. Current solutions either compromise privacy by storing your data in the cloud or are restricted to individual tools like ChatGPT's current memory implementation.

We were prototyping this solution with other models, but the release of gpt-oss models presented a breakthrough opportunity. For the first time, we have open reasoning models powerful enough to understand complex patterns in how people work, yet lightweight enough to run locally—preserving privacy while delivering intelligent context. This hackathon became our chance to solve the context fragmentation problem at its root: by creating a memory layer that works with the AI tools you already use, respects your privacy, and actually understands how you work. The potential impact is enormous—not just for individual productivity, but for making AI assistants truly personal and effective partners in our daily work.

What it does

Our browser memory extension adds an invisible memory layer to your browser that automatically compresses your browsing activity into a portable context for any AI agent. The key hackathon contribution was integrating gpt-oss-20b to power this memory layer with advanced reasoning capabilities. When using ChatGPT, Claude, or other AI tools, the extension provides a context button that inserts relevant information from your browsing history—enhanced by gpt-oss's ability to identify interesting URLs and fetch their content through specialized tool calls.

How we built it

Our browser memory extension began as a functional tool designed to maintain context across AI conversations by analyzing browsing patterns. The core architecture was built with a privacy-first approach, storing all user data locally in Chrome's storage system and providing a side panel interface for context management.

For this hackathon, we focused specifically on enhancing our existing foundation with gpt-oss integration. Our key contributions were:

• Implementing seamless integration with OpenAI-compatible API endpoints to run gpt-oss-20b locally via Ollama, LM Studio, or vLLM
• Developing custom tool calls that allow gpt-oss to fetch content from URLs it identifies as relevant in your browsing history
• Optimizing the model to work within browser constraints while maintaining the reasoning capabilities that make gpt-oss unique

Rather than just analyzing browsing history, gpt-oss-20b actively reasons about relationships between pages and intelligently fetches additional context when needed. This builds upon our original architecture while adding the sophisticated reasoning that only gpt-oss can provide—turning passive data into actionable AI context without compromising privacy.

Challenges we ran into

Integrating gpt-oss-20b into our existing extension presented unique challenges:

• Implementing a reliable tool call for URL content fetching
• Optimizing gpt-oss-20b to run efficiently within constraints without sacrificing reasoning quality

Accomplishments that we're proud of

• Successfully integrated gpt-oss-20b into our existing browser extension, creating the first gpt-oss-powered memory layer for AI context in the browser
• Developed specialized tool calls that enable gpt-oss to fetch relevant page content based on URL analysis
• Demonstrated gpt-oss-20b's unique value for understanding complex temporal patterns in user behavior
• Published the enhanced version to the Chrome Web Store

What we learned

• gpt-oss-20b's reasoning depth is essential for understanding meaningful connections between browsing sessions
• Tool-calling capabilities transform how context can be dynamically retrieved and applied
• The existing architecture of our extension provided the perfect foundation to showcase gpt-oss's unique strengths
• Local processing with gpt-oss creates superior personalization compared to cloud-based alternatives
• Privacy and advanced functionality can coexist when leveraging open models on local hardware

What's next for Browser Memory Extension

• Expand gpt-oss integration to include more sophisticated workflow pattern recognition
• Develop hardware-specific optimizations for even better performance with gpt-oss models
• Create adaptive learning that anticipates needs based on work patterns using gpt-oss reasoning
• Build MCP server integration to expose browsing patterns as a standardized context protocol
• Optimize for lower-end hardware to bring gpt-oss-powered contextual AI to developing regions

Our hackathon work transformed our browser memory extension from a simple context manager into an intelligent memory layer powered by gpt-oss's unique reasoning capabilities. This integration demonstrates how open models can create truly personal AI experiences that understand how you work—without compromising your privacy.

Built With

• biome
• browser-extension
• lm-studio
• ollama
• svelte
• typescript
• vllm