Unbury: An Open-Source AI Memory That Turns Messy Life Into Things You Can Actually Do

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

#

What I Built

I built Unbury for a close friend who has a habit many of us share: important things get buried in everyday conversations.

A deadline might be mentioned in a fast-moving WhatsApp chat. A task might be hidden inside a screenshot. Someone sends an important PDF or a voice note containing something they need to remember days later. The problem isn't that they don't care about these details—the problem is that the information is scattered everywhere across multiple channels.

I didn't want to build another standard todo app where they had to manually convert everything into tasks. I wanted to build something that could understand raw, messy inputs as they arrive and automatically extract actionable commitments and useful memories.

That's how Unbury started. Unbury is an AI-powered external memory for messy real-life information. You can dump text, screenshots, PDFs, or voice notes into it. It parses the content, extracts potential tasks and useful memories, and presents a confirmation review before anything is committed to memory.

The core loop is:

Dump → Understand → Review → Confirm → Remember → Remind

For example: "Please send the one month strategy to Rahul tomorrow at 5 PM and remind me 15 minutes before."

Transforms into:

Task: Send the one month strategy to Rahul #

Deadline: Tomorrow at 5:00 PM #

Reminder: 15 minutes before #

Evidence: The original raw message

If information isn't actionable (like casual chatter or background thoughts), Unbury filters it into memories rather than unnecessary tasks, keeping the action feed clean.

#

Demo

The deployed application uses hosted Gemma through Google's API for production inference. You can test the core workflow:

1. Dump unstructured text, upload a document, or share notes.
2. Let Unbury parse the content.
3. Review extracted tasks, deadlines, and memories.
4. Edit, adjust, or discard irrelevant suggestions.
5. Confirm what gets committed to memory.
6. Query stored data through the grounded Ask assistant.

#

Code

Turn messy information into structured tasks and memory — without losing context.

Unbury is an AI assistant that reads whatever you throw at it — messages, PDFs, voice notes, screenshots — and extracts the tasks, deadlines, and durable facts buried inside. Nothing is saved without your review.

Live Demo #

unbury.onrender.com

What It Does #

Dump anything — paste text, upload a PDF, drop a screenshot, or record a voice note #

AI extraction — Gemma identifies tasks with deadlines and facts worth remembering #

Review before saving — every extraction is a proposal you confirm, edit, or discard #

Progressive reminders — browser notifications fire at 1 week, 3 days, 1 day, and 1 hour before a deadline (while the app is open and permission is granted) #

Persistent memory — durable facts (people, roles, context) are stored separately from tasks and surface when relevant #

Ask Unbury — query your saved memory in plain…

The repository includes the complete full-stack implementation:

• AI extraction pipeline and prompt parsing
• Multimodal ingestion (OCR, PDF parsing, text transcripts)
• Structured review and user confirmation engine
• Reminder scheduler and grounded retrieval workflow

#

How I Built It

Unbury is architected around Gemma as its cognitive core:

Local Development: Gemma running locally viaOllama for low-latency iteration. #

Production Deployment: The Next.js application and AI API routes are deployed onRender , using hosted Gemma through Google's API. #

**Data Persistence:**MongoDB Atlas database cluster storing structured tasks, memories, review states, and reminder queues.

Architecture Flow:

#

Why Does Open Innovation Matter?

Open-weight models like Gemma are essential for personal productivity and memory systems for two reasons:

Privacy & Data Sovereignty: Everyday life dumps contain sensitive personal data—chats, financial deadlines, screenshots, and internal notes. With open-weight models, users are not forced to send their personal life data to proprietary third-party LLMs with ambiguous retention policies. The exact same pipeline can run completely offline and locally using Ollama. 2. Predictable Scalability: Open architectures eliminate arbitrary token rate limits, sudden API deprecations, and prohibitive subscription fees, enabling sustainable personal tools that remain accessible to everyone.

#

Prize Categories

Best Use of Gemma (Gemma integrated via Ollama for local environments and hosted Gemma for the live production pipeline) #

Best Use of Render (Full-stack web application, inference routes, and background scheduler deployed on Render) #

Best Use of MongoDB Atlas (Managed MongoDB Atlas cluster handling persistent memories, review states, tasks, and scheduling metadata)