Harsh Wardhan
Identity: Semantic SearchHackathonSemi-FinalistOpen Source

MemoryOS

AI-Powered Semantic Second Brain & Note Retrieval Engine

Engineering Challenge

“How do you build an AI second brain that understands context, not keywords?”

AI•Completed•Full Stack Builder•36 Hours (SRM Builds 7.0)
Build Time36 HoursSRM Builds 7.0
HackathonSemi-FinalistTop Project
EmbeddingsVectorOpenAI API
Search Latency< 200msSemantic Query
MemoryOS
01 Product Context

Problem & Product Goals

Traditional note-taking tools fail when users forget the specific keywords used when writing a note. MemoryOS leverages OpenAI vector embeddings stored in Supabase to provide real-time semantic retrieval, allowing users to query their notes naturally.

Why I Built It

During study sessions, finding past notes often required manual searching through folders or trying multiple keyword combinations. I wanted a system that understood the semantic context of what I was looking for.

02 Implementation

Technical Architecture & Core Engine

Design Philosophy: Focus on a single core workflow—write markdown, auto-generate embeddings in the background, and search by meaning—avoiding feature bloat during a tight hackathon timeline.

03 Technical Rigor

Engineering Decisions

Decision #01 • Background Embedding Queue vs Synchronous API Calls
Problem:

Generating embeddings on every keystroke blocked the main thread and caused visible editor lag.

Decision:

Implemented a debounced background queue that sends note contents to OpenAI embeddings endpoints only after editing pauses.

Tradeoff:

Brief delay (500ms) before newly edited text is searchable via semantic queries.

Outcome:

Preserved smooth 60fps markdown typing performance while maintaining search indexing.

04 Problem Solving

Challenges & Solutions

36-Hour Hackathon Scope Management

Issue: Attempting to build rich text tools and complex folder hierarchies threatened completing the core semantic search flow.

Decision/Solution: Cut all peripheral features to focus exclusively on perfecting the write -> auto-embed -> semantic query pipeline.

Future Roadmap

What's Next

Local embedding generation using WebGPU models
Graph visualization of connected note concepts
Encrypted local notes storage
Retrospective Summary

Key Takeaways

Takeaway #01

Scoping down to one exceptionally well-executed feature wins over long lists of half-built capabilities.

Takeaway #02

Vector search changes how personal information management tools can be structured.

Continue Exploring

Related Engineering Projects