Verath: AI-Powered Personal Memory System
Vishakha Sanjay Yadav
Verath
AI-Powered Personal Memory System
Version 3.0.0 · Record · Extract · Search · Remember
Overview
Verath is an intelligent personal memory system. You record voice notes, meetings, or thoughts; the backend transcribes audio, extracts structured metadata (intent, entities, dates, importance), stores everything in MongoDB (documents) and ChromaDB (vectors), and lets you query your history in natural language using retrieval-augmented generation (RAG) with cross-encoder re-ranking.
Typical user journey
Capture — Speak into the mobile app, use the web dashboard, hit
POST /record, or run the CLI. Process — Whisper transcribes locally; the memory extractor cleans text, detects corrections, classifies intent, and scores importance.
Store — Each memory gets a UUID, Gemini embedding, MongoDB document, and ChromaDB vector (per-user collection).
Recall — Ask “What meetings did I have this week?”; the query engine retrieves 20 candidates, re-ranks to top 5, and Groq/Gemini answers from your data only.
Act — Reminders fire for dated meetings/deadlines; export or delete memories; view timeline, graph, and insights.
Features
Memory & intelligence (v3.0)
Feature | Description |
|---|---|
Speech correction detection | Parses self-corrections in transcripts |
Intent classification | meeting, deadline, task, commitment, reminder, etc. |
Entity extraction | Dates, people, locations, organizations via NLP + dateparser |
Importance scoring | Intent-based boosts for deadlines and meetings |
Hybrid RAG query | Vector retrieval → cross-encoder re-rank → LLM answer |
Paginated timeline | Today’s memories with page / page_size |
Daily summary & insights | LLM-generated briefings and patterns |
Memory graph | Nodes and edges for visualization ( GET /graph) |
Export | JSON or CSV with optional intent and date filters |
Delete memory | Removes from MongoDB and ChromaDB atomically |
Platform & operations
Feature | Description |
|---|---|
JWT authentication | Signup, login, refresh rotation, logout blacklist |
Rate limiting | Signup 5/min, login 10/min, refresh 20/min (per IP) |
WebSocket | /ws/updates?token= for live updates and ping/pong |
Background worker | Recording jobs, compression, retries (2s → 4s → 8s) |
Dead-letter queue | Failed tasks inspectable and manually retried |
Tech stack
Category | Technology |
|---|---|
Framework | FastAPI, Uvicorn |
Validation | Pydantic 2.x, pydantic-settings |
Speech | faster-whisper |
LLM | Groq (default llama-3.1-8b-instant, fallback llama-3.3-70b-versatile in queries) |
Embeddings | Google Gemini text-embedding-004 |
Vector DB | ChromaDB 0.5+ (HNSW, cosine) |
Document DB | MongoDB via Motor (async) |
Re-ranking | sentence-transformers cross-encoder |
Scheduler | APScheduler (reminders) |
Auth | python-jose, bcrypt, slowapi |
Audio I/O | sounddevice |
System diagram
Project structure
Verath/
├── backend/
│ ├── app/
│ ├── tests/
├── web/
├── mobile/
├── scripts/
└── README.md
Like this project
Posted Jun 15, 2026
A personal memory system for recording and recalling thoughts, meetings, and notes.
Likes
0
Views
0
