Where your scattered day becomes one memory.
Sixteen tabs, four meetings, a hundred half-read threads — you saw it all today and can recall almost none of it. Yogurt is the OS for the mind I designed and built to fix that: on-device, ambient, always listening to my day so I don’t have to.
Read the deep-dive(01) — The Catalyst
What if my operating system could remember everything I saw?
As a product and UX designer building complex B2B AI SaaS platforms, my daily cognitive load became unsustainable. I was constantly jumping between Figma, Slack, Zoom, and Chrome. The context was always there — but it was siloed.
“AI is fundamentally a reasoning engine, but its bottleneck is personal context.”
— Satya Nadella
He was right. We spend entirely too much of our day acting as human routers — searching our own history rather than acting on it. Yogurt was born from a simple question: what if my OS could remember, so I didn’t have to?
(02) — The Scale of the Problem
Context-loss compounds at every level.
16h
Personal
We waste hours of our 16-hour daily attention span trying to recall decisions made in passing.
Silos
Team
Knowledge locked inside individual apps — manual status updates and standups just to stay level.
IP
Enterprise
When an employee leaves, their context leaves with them. Organizational memory isn't captured at the OS level.
(03) — The Paradigm Shift
The future of AI isn’t a chatbot in a browser tab.
The future is Local Ambient Memory. Yogurt acts as an OS for your mind — software that sees what you see, hears what you hear, and securely indexes it 100% on-device. It turns raw screen time into a queryable, encrypted database — without your data ever hitting a cloud server.
→ The real Rewind window: every day of history, grouped into sessions, browsable offline. This is a working HTML replica of the shipped app — not a screenshot.
Rewind
356 memories · Today
🔒 on-device history
Switched to Google Chrome…
[Meeting action items] Meeting in Google Chrome ended…
Building Synapse: Local-First Memory — Google Gemini…
Pragathi Netha | LinkedIn — High memory usage — 1.2 G…
(04) — Yogurt in Action
Four surfaces, one memory.
Text-mode replay, not a video vault.
A scrubbable heatmap of the day that replays the exact text on your screen at any given second.
Why — Competitors record heavy, battery-draining video — unsearchable and slow. Yogurt captures text via Vision OCR and Whisper: mathematically efficient, instantly queryable, stripped of UI noise.
Google Chrome — Product Designer: AI-B2B & Enterprise
10:39 PMThis is a showcase of what happens when curiosity drives the process. AI-native helpdesk platform for managing complex customer support workflows and decision-making.
› Screen Elements 16
A librarian who works while you don't.
An auto-generated, tagged wiki of your memory cards.
Why — Raw OCR is a server log; human memory needs structure. A background local LLM distills screen chaos into durable facts — decisions, commitments, preferences — with zero folders organized.
Synthesized Memories
Facts, decisions, and preferences locally extracted.
Yogurt CASE STUDY was discussed as a local-first ambient memory layer for macOS.
WHOOP iOS app's features and functionality were analyzed via Mobbin.
Ramachandran's design process was analyzed as a case-study reference.
⌃⌥ — and the AI meets you at your cursor.
A global hotkey summons a context-aware AI exactly where your cursor rests — inside a LinkedIn message box, mid-thread.
Why — Context switching destroys deep work. Via Apple's Accessibility APIs, Yogurt meets you inside your active input field, fully aware of the screen, drafting replies in your voice.
Pragathi Netha
Hey Parth! Loved the Yogurt teaser. Could you show me how the enterprise dashboard handles memory at scale?
▍
context: LinkedIn thread + 3 memories · drafted on-device
Pipes: memory that acts while you sleep.
Zero-code automations that route ambient data to Notion, Linear, and Slack.
Why — Passive recall is half the battle; utility comes from action. Yogurt syncs CRM data, drafts executive briefs, and updates Jira tickets overnight.
The
Wednesday Brief
Hello Parth, I’d like to discuss today’s work on Building Synapse: Local-First Memory.
on-device · llama3.2
Push your work forward
Let’s do it →Resume Building Synapse: Local-First Memory
A productive day — 120 screen reads across 2 apps, plus a review of UX/UI interview questions to sharpen agent workflows.
Top to-dos
◯ Continue screenpipe: Workflow Memory for AI Agents
◯ Continue [Meeting action items] follow-ups
Ask it something. Watch it think.
This is the actual retrieval-and-synthesis loop, not a video. Pick a question — the agent types it, streams its tool calls as it searches your local memory, then composes the answer.
›Ask your team's memory…
(05) — Engineering Challenges
Building a localized brain wasn’t without hurdles.
The Reader Heuristic
Vision OCR returns everything — buttons, nav bars, the word “Hide” forty times. We engineered a Reader Heuristic to filter OCR noise: ignoring UI fragments and stitching real sentences back together across broken bounding boxes, so the replay reads like a page, not a log.
Accessibility focus traps
Apple's Accessibility API has strict focus semantics, and Chromium browsers routinely refuse to expose the active text field. The Proactive Copilot ships a tiered fallback ladder — exact caret bounds, then the browser window frame, then the pointer — so the hotkey always answers.
The watermark system
A local LLM chewing through your screen all day is a battery argument waiting to happen. An active watermark system ensures the model only processes new captures — each exactly once, batched into a single pass — for zero impact on MacBook battery during heavy meetings.
(06) — The Horizon
V1 is perfect recall.
V2 is autonomous agency.
We are moving toward a reality where your local memory node can federate with your team’s nodes — a secure “Company Brain” that eliminates standup meetings entirely, while retaining strict hardware-locked privacy boundaries for the individual. Your team learns what you shipped. It never learns what you saw.
Already shipping toward it
The rules engine that will drive Pipes exists today. Watch it assemble the seed automation — meeting ends, action items extracted, memory written, notification fired — the same loop that will one day update your CRM instead.
New Rule
IF
THEN
● Ran 16× · last Jul 15, 2026 at 7:20 PM — Extracted action items · Saved to memory · Notified
The thesis, in his words
(07) — The Outcome
From a designer’s friction to a working brain.
Every feature above shipped as working SwiftUI — a tested rules engine, a fact synthesizer, a copilot that survives real browsers. The deepest lesson: design and engineering aren’t phases. The Reader Heuristic is a typography decision enforced in Swift; the watermark is an ethics decision enforced in battery life.
100% on-device · AES-GCM at rest · opt-in by architecture