Friend

• 

  Chatbot with memory for no context switching (Friend!)

• 

  Ghost suggestion text ( Tab to complete)

• 

  Saved memory ( add/ remove )

• 

  User profile

Inspiration

Think about: • That Airline confirmation number you copied weeks ago but can’t find when you’re checking into your flight. • The Zoom link buried in chat history right before a meeting. • A GitHub URL you used for an application last month. We wanted to fix that by making memory retrieval frictionless - like a 'friend' who knows exactly what you want to type and when you want to type it.

What it does

    •    Friend remembers things you’ve saved ( select any text - right click - add to memory )     •    It retrieves the right context instantly based on the current page and the text box using DOM.     •    Example: autofill your email when signing up, paste your GitHub when applying for jobs, recall a LinkedIn bio for outreach, Bring up your saved items from reminder when you're shopping, suggesting your airplane number when you try to track it 3 months after booking and anything you want to fill in any text box inside your browser ( for now )     •    It’s like supercharged paste memory-first, context-aware, and lightning fast and fully private and local - You can save anything and it's all locally in your device.

How we built it

• Local-first setup: Powered by Ollama or LM Studio running GPT-OSS 20B model.
• Mini agentic search: Friend runs a lightweight agentic search and instantly finds the most relevant memory based on the context.
• Extension only: No server, no database - everything lives in your browser and local. sub second retrieval , private, and self-contained. No data leaks whatsoever. Period.
    •.    Context Awareness - we parse the entire DOM of any website based on your current cursor location and that gives our the right contextual information for the retrieval.
    •.     Supports more than 200 memory entries seamlessly ( and beyond probably. Give it a try!)    

Challenges we ran into

    •    Balancing speed vs accuracy in retrieval (fast enough to feel magical, smart enough not to annoy).     •    Normalizing DOM inputs across different sites. ( What is relevant for the current task?)     •    Deciding what memory to keep, and how to give users control over it (right-click to save/remove).     •   designing Intuitive UX ( Auto suggestion, sidebar chat panel when needed )

Accomplishments that we're proud of

    •    Built a working memory-first 'friend' layer that retrieves the right info in under 200ms.     •    Demoed cross-site usage: job applications, forms, email replies - not just dev/code use cases.     •    Clean UI and UX     •    Early feedback: people said it “feels like magic.”. Fully shippable and usable that is fully private! (Waiting to upload to chrome store for anyone to use it if this is a successful hack ;) )

What we learned

• Memory is more powerful than history → Logging is easy, but the value comes from surfacing the right memory at the right time.
• Context is the real challenge → Storing snippets is simple; knowing when and where to bring them back is much harder.
• Trust is everything → The hardest problem isn’t technical. Users want control over what’s remembered and when it’s recalled.

What's next for Friend

• Smarter context signals

Make Friend better at knowing when to surface a memory (not just DOM fields, but intent and workflow cues). • User controls & transparency Let people see, edit, and manage exactly what’s remembered so trust stays high. • Agentic control Move beyond recall — let Friend take actions directly in the browser (click buttons, scroll, fill forms), reducing friction even further. • Cross-device memory Sync snippets securely so Friend works across laptop, phone, and tablet. • Tiny magic moments Polish the UX so retrieval feels less like “searching” and more like memory just appearing when you need it.

Built With

• gpt
• lmstudio
• nvidia
• ollama