🧠

My Brain

Not another to-do app.
A second brain that listens to your day, remembers your life, and mentors you back.

Designed by Sandeep Karnati

0
minutes of effort a day
0
days woven into one book a year
0%
yours — private & self-hosted

The problem

Every app organizes your tasks.
None of them understand your life.

A normal to-do app

  • ☑️ Stores what you type
  • 😶 Forgets how the day felt
  • 📦 A filing cabinet — tasks go in, silence comes out
  • 🔁 Day 1000 works exactly like day 1

My Brain

  • 🎙️ You speak, it builds the day
  • 📖 Every evening becomes a written story
  • 🧭 A mentor reads it all and answers back
  • 📈 Day 1000 knows 999 days more about you

How a day works

One day in the loop — tap each step

🌅
Morning dump
60 seconds
Your day, built
AI-crafted tasks
🌙
Evening story
your words
📖
Story of the Day
auto-written
🧭
Mentor speaks
honest, nightly

👆 Tap a step to see what happens there.

One story · two lives

It separates work from family — and watches the balance

The client call you dreaded turned into the win of the week — they signed. You closed three threads you’d been circling for days, and it shows: you ended work lighter than you started it.

Work mood
Home mood

🧭 Mentor: “Work is winning at 78 while home sits at 52 for the third week. The proposal can wait one evening — Thursday’s dinner cannot.”

Feel it

60 seconds of speaking → a day that runs itself

🎙️ 7:02 AM — you, speaking:

✨ your day, one second later:

📌 Send the revised quote to the architect [💼 work · high · 30m]

○ Book the dentist for Amma [🏠 personal]

○ Pick up the school project supplies [🏠 personal]

○ Review the vendor contract [💼 work · 45m]

Chapter 2

One brain, ten rooms

Tasks are just the heartbeat. Tap any room to see what lives inside.

👆 Tap a room.

Everything in every room flows into the same memory — that’s the next chapter.

Chapter 3

The architecture, as a living picture

Four layers. Everything you see talks to the layer below it — and your data never leaves your control. Tap any block.

📱
Phone app
installable · works like a native app
✈️
Telegram companion
voice notes · reminders · save anything
🖥️
Your private server
the app + its engines: the story-weaver, the mentor, the watchdogs — ticking on schedule all night
🗄️
App database
tasks · stories · settings
📚
Memory vault
meaning-searchable, on your server
☁️
SuperMemory
second memory, in the cloud
🤖
AI brains
called only for moments of intelligence — they keep nothing
🏠
Home backup
verified copies, every night

👆 Tap any block of the architecture.

For the curious — the real stack
A React progressive web app and a NestJS (TypeScript) API, shipped as one Docker image behind a Caddy reverse-proxy with automatic HTTPS. Operational data in SQLite via Prisma; the memory vault is a Postgres database with the pgvector extension running beside the app; SuperMemory is the cloud twin. Background engines are in-process schedulers (per-minute ticks) for the nightly story, mentor, weekly review, monthly chapters, suggestions and watchdogs. Nightly backups: SQLite online snapshot + Postgres dump, pulled (not pushed) by a home server over SSH, with checksums and integrity checks on every single pull.

Chapter 4

How remembering actually works

This is the heart of the whole system — and the part most apps don’t have at all.

Writing a memory — what happens when you save anything

📄
Anything arrives
a document · a link · a video · tonight’s story
✍️
AI distills it
a faithful summary — videos are literally watched
🧬
Meaning, not words
converted into a “meaning fingerprint”
📚☁️
Written twice
memory vault + SuperMemory — a queue retries until BOTH have it
For the curious — the write path
Content → model-generated summary → OpenAI text-embedding vectors → an outbox queue writes to both stores (RAG/pgvector + SuperMemory) with retries, so the two memories can never silently drift apart. Every memory carries tags (bookmark, idea, activity, skill…) that power the scope system below. Long documents are auto-chunked along headings with overlapping windows so retrieval can land on the exact passage.

Finding a memory — ask, and watch what happens

This is a live simulation — tap a question.

🔖 Video: “Technical SEO in 2026”
🌙 Story · the launch evening
📄 Doc: pricing research
🔖 Article: keyword clustering
💡 Idea: newsletter engine
🌙 Story · the family Sunday
🔖 Guide: local SEO checklist
📄 Doc: vendor comparison
For the curious — the read path
Your question is converted into the same kind of meaning-fingerprint and compared against every memory — matching by meaning, not keywords (“cloud SEO skills” finds a video that never says those words). The chosen scope (Bookmarks, Capture, My life…) filters by tags on BOTH stores, results are merged and ranked, and the answer is grounded in the retrieved passages with sources attached. If the scope truly has nothing, it says so honestly instead of guessing.

📚 The memory vault (on your server)

Stores everything verbatim — the exact words, retrievable exactly. Built for receipts: when the app answers, it can show you the original.

☁️ SuperMemory (cloud twin)

A second, independent memory with its own intelligence — redundancy by design. Two memories, one truth: the queue keeps them in lock-step.

Chapter 5

Talk to your brain — with receipts

Every conversation is scoped to a part of your life, and every claim comes with its source. Watch a real exchange replay:

🌐 Everything 📊 My life 🔖 Bookmarks
When did I last feel really great?
🧭 deciding: this needs a memory search → searching My life
🌙 Story · Jun 8 🌙 Story · May 30 📜 Weekly review

🎯 Scoped, not soup

Ask your bookmarks, your documents, your ideas, your skills — or your life. Each scope searches only its own shelf, and says so honestly when the shelf is empty.

🧾 Receipts attached

Answers cite the actual saved items they came from — tap a source and read the original. No hallucinated confidence.

⚡ A router thinks first

A fast small model decides per message: does this need a fresh memory search, or is it a follow-up? Searches happen only when they should.

For the curious — chat internals
Sessions carry a scope; a lightweight router model classifies each message (new retrieval vs follow-up) and writes a standalone search query when needed. Retrieval merges both memory stores with tag filters, then the answer streams token-by-token to the screen. A per-document chat mode binds a session to one document. Conversations are kept on a retention schedule you control.

Chapter 6

The right brain for the right job

One AI model would be either wasteful or weak. My Brain runs a crew — each picked for its task, each swappable from Settings without touching code.

🖋️

The Storyteller — a deep, careful model

Writes what deserves craft: the nightly Story of the Day, the mentor’s guidance, the Sunday review, the monthly chapters, the book of the year.

The Sprinter — a fast, tiny model

Handles hundreds of small jobs: routing chat messages, cleaning voice transcripts, spotting people’s names, phrasing the 4 PM nudge. Instant and frugal.

👁️

The Watcher — a model with eyes

Literally watches the YouTube videos you bookmark and reads the pages you save, writing faithful summaries so your memory holds the substance, not just the link.

👂

The Ears — speech-to-text engines

Turn your voice into clean text everywhere — the morning dump, the night story, Telegram voice notes — with a choice of engines and automatic fallback.

🎛️ Every role has its own picker in Settings — and every instruction the AI follows is an editable prompt. The owner can re-tune any brain, any time, without a developer.

For the curious — the actual crew
Claude Sonnet for the long-form writing and mentoring; Claude Haiku for routing, cleanup, extraction and nudges; Gemini (multimodal) for watching bookmarked videos and reading pages; GPT-4o-transcribe / Whisper, ElevenLabs Scribe and Deepgram Nova as selectable speech engines. All routed through one gateway with per-request usage logging, and every role’s model + prompt is user-configurable.

Chapter 7

Connected to your world, guarded like a vault

Tap a connection to see its role. Every key is encrypted at rest — even someone holding the database file couldn’t read them.

👆 Tap a connection.

For the curious — the key vault
All third-party credentials live in an encrypted connector registry (AES-256-GCM, key derived from a server-side secret) — never in code, never in the repository. Each connector has a live “Test” button in Settings that makes one real, cheap call to prove the key works.

Chapter 8

Built by AI — held to engineering discipline

The entire application was built by an AI working under an enforced pipeline — nothing ships on trust.

🎫
Ticket
every feature is tracked
⌨️
Code
one branch per feature
🧪
130+ tests
red = nothing ships
🚀
Deploy
all-or-nothing gate
Verified live
then — and only then — done

Everything versioned on a private repository · data backed up to a second machine nightly, with every backup integrity-tested.

From zero to a life companion — in days, not years

Born

Skeleton, login, the twin-memory engine, the design system.

It started remembering

Capture for documents; bookmarks imported and literally watched; ideas and skills organized by AI.

It started listening

The daily loop: morning dumps, evening stories, the woven Story of the Day, Telegram by voice.

It started understanding

The mentor, focus areas, the on-track score, suggested tomorrows.

It grew a long memory

Weekly reviews with patterns and experiments; monthly chapters; the Story of the Year; people memory.

It embraced the whole life

Work and family as equal citizens — two narratives, two moods, one mentor watching the balance.

Most apps manage what you do.
This one understands who you are.

Every day it runs, it knows one day more. The version you meet next year will know its owner better than any product ever has.

🧠 My Brain — one life, fully remembered