i.
Import
Export your chat history from any platform. Drop the files in. Pratibmb auto-detects the format and parses everything into a local SQLite database.
A local-first memory machine · v0.5
A letter to
who you used to be.
Import a decade of WhatsApp, iMessage, Instagram, Gmail, Telegram, Twitter, Discord, or Facebook — and a model that runs entirely on your laptop learns the cadence of your old self, and lets you text them back.
Somewhere on your phone, in a folder you haven't opened in years, there is a person who no longer exists. They're still making jokes you wouldn't make now. Still using emojis you'd be embarrassed by. Still in love, or afraid, or ambitious about things you've long since resolved or forgotten.
Pratibmb takes those thousands of real messages — the ones you actually sent and received — and builds a small local model that talks the way you used to talk. Not a persona. Not a chatbot. A close-enough sketch of a self you can greet again.
And because the one dataset you should never hand to somebody else's API is a decade of private messages, every byte stays on your machine.
"I found messages from 2014 where I was excited about a job I'd completely forgotten about."
— early tester
"My 2016 self texted completely differently. More emojis, more slang, more exclamation marks."
— early tester
"Chatting with past-me about my old friendships was unexpectedly moving."
— early tester
Four steps, no servers.
Architecture · fig. 01
ii.
Embed
A tiny 84MB embedding model runs on your CPU or GPU and builds semantic vectors of every message. Not a single byte leaves your machine.
iii.
Profile
A 2.3GB local LLM extracts the shape of your past: the people who mattered, the events, the interests, the verbal tics and cadences that made you, you.
iv.
Chat
Pick a year. Start typing. A younger version of you replies — in your voice, grounded in your real memories. Optional LoRA tuning makes it eerier.
Eight platforms, one format.
Supported imports
W
WhatsApp
Export chat · .txt
F
Facebook
Download Your Information · JSON
I
Instagram
Download Your Information · JSON
G
Gmail
Google Takeout · MBOX
iM
iMessage
macOS chat.db · SQLite
T
Telegram
Desktop export · JSON
X
Twitter / X
Archive · .js
D
Discord
DiscordChatExporter · JSON
Privacy isn't a policy. It's the architecture.
Receipts, not promises
01 · Local
On your hardware
Every computation — embedding, profiling, inference, tuning — runs on your own CPU or GPU. Messages never leave the machine.
02 · No cloud
No API keys
No OpenAI, no remote inference, no silent upload. Works with Wi-Fi disabled after the first model download.
03 · No telemetry
No phone home
Zero analytics, zero crash reports, zero usage tracking. Inspect the network. Inspect the source. Don't trust — verify.
04 · Open
Auditable
AGPLv3 for personal and open-source use, with a commercial license for everything else. Every line of code is readable.
Install in one line.
Free · Open source · ~2.5 GB one-time
One-command install
# macOS / Linux
curl -fsSL https://pratibmb.com/install.sh | bash
# Windows PowerShell
irm https://pratibmb.com/install.ps1 | iex
macOS · Apple Silicon + Intel
macOS
M1 / M2 / M3 / M4 and Intel. MLX fine-tuning included out of the box on Apple Silicon.
brew tap tapaskar/pratibmb brew install --cask pratibmb
Linux · Ubuntu / Debian / Arch
Linux
.deb, .AppImage, or AUR. CUDA acceleration for fine-tuning on NVIDIA; CPU fallback available.
yay -S pratibmb-bin
Windows · 10 / 11 · 64-bit
Windows
Winget, MSI or .exe installer. NVIDIA CUDA acceleration for fine-tuning on supported GPUs.
winget install tapaskar.Pratibmb
Requires Python 3.9+ and ~8 GB RAM. Models (~2.5 GB) download on first launch.
Optional: learn your voice.
LoRA · adapters · local
The base model already captures the bulk of your communication style. LoRA tuning sharpens the details — your abbreviations, your code-switches, your emoji habits.
Platform
Backend
GPU
Time ≈ 1500 pairs
Install
macOS · Apple Silicon
MLX-LM
Metal (no GPU)
~20 min
Included
Windows / Linux · NVIDIA
PyTorch · PEFT · QLoRA
6 GB+ VRAM
~30 min
pip install "pratibmb[finetune-pytorch]"Linux · CPU-only
PyTorch
None
~2 hours
pip install "pratibmb[finetune-pytorch]"macOS · Intel
—
—
—
Not available
Pratibmb · Clinical — in preview
The patient's history is already on their phone.
A clinical extension for licensed mental health professionals: sentiment timelines, communication-pattern analysis, and patient-controlled session-prep briefs. Same local-first architecture. Designed to surface patterns for clinician review — never to diagnose.
Questions, asked quietly.
FAQ
Is this therapy?
No. Pratibmb is a conversation tool, not a mental-health product. If you use it to process grief or difficult emotions, please set time limits and talk to a professional. Built-in session limits and gentle reframes are on the v1.0 roadmap.
Can anyone else see my data?
No. Everything runs on your hardware. There are no servers, no accounts, no cloud storage, no analytics. Your messages, embeddings, models, and profiles stay on your machine. Uninstall the app and the data is gone.
Does it work offline?
Yes. After the one-time model download (~2.5 GB on first launch), Pratibmb works with no internet connection. It uses a local LLM (Gemma-3-4B) and a local embedding model (Nomic Embed). No API calls, no cloud inference.
How accurate is it?
The base model captures your general style, topics, and relationships well. Optional LoRA fine-tuning (~20 min on Apple Silicon) teaches it your specific texting patterns — abbreviations, language mix, emoji habits. Not perfect, but often surprisingly close.
What hardware do I need?
Any modern computer with 8 GB+ RAM. Apple Silicon Macs get the best experience (Metal GPU acceleration). NVIDIA GPUs accelerate fine-tuning on Windows/Linux. CPU-only works too — just slower.
Why not just use ChatGPT with my messages?
Because your private messages are the most intimate data you own. Uploading a decade of conversations to any cloud API means that data now lives on someone else's servers, subject to their terms, their breaches, and their training pipelines. Pratibmb exists so you never have to make that trade-off.
macOS says "damaged and can't be opened"
The app isn't signed with an Apple Developer certificate (code signing costs $99/year). It's safe — you can verify the source yourself. To fix, run in Terminal: xattr -cr /Applications/Pratibmb.app
Can I use this commercially?
Not without a license. Pratibmb is AGPLv3, so personal use, research, and open-source projects are free. Embedding in a proprietary product, offering as a hosted service, or distributing in closed-source software requires a separate commercial license. Contact admin@sparkupcloud.com.
Can I use it for a deceased family member's messages?
Technically, yes — you can import their exported messages and chat with the model. This is emotionally powerful and should be approached with care. We recommend session time limits and remembering that it is a model, not the person. A dedicated "legacy mode" with additional safeguards is planned for v1.0.
What would you say
to 2014·you?
Free. Open source. Runs offline after first launch. Your messages stay where they belong — on your machine.