A journal that remembers, built for neurodivergent adults. Speak or type what is going on, on the iPhone or straight into the Apple Watch, and Cortex keeps every word, sorts it, and lets you ask an assistant that answers from your own entries. It never tells you what caused what. Your journal lives on your device, and when AI features need processing, the relevant text (with direct identifiers redacted) travels encrypted to AWS Bedrock and is not retained by the AI provider.
A day with Cortex
Hold the button and talk, on the phone or on the Watch. The transcript knows your medication names. Nothing to pick first, no form to fill, no streak to keep.
Every entry gets its type, its subjects and, when it fits, a thread. One short line comes back so you know it landed. Weeks later you find it again without ever having tagged anything.
Ask Mimo what you noticed the last time you changed a dose. Open a thread to read everything about one question, in order. Write it up for the appointment in a tap.
Speak naturally. On-device transcription understands medication names, emotional shorthand, and neurodivergent speech patterns that standard dictation gets wrong. Or type, whatever takes less energy.
Say an entry into the Apple Watch and it lands in your journal, turned into text on the watch itself. Today's medications are on the wrist too, and each dose logs with one tap. No phone to find, nothing to unlock.
A thread is a question you are working on, like "why do my teeth hurt?" or "why do outings drain me?". File entries into it, or let Cortex suggest them, and the thread page shows everything you ever wrote about it in order.
Medications, supplements, and accommodations organized by time of day. Log doses with a tap on the phone or the Apple Watch, track what you've taken, and see your routine at a glance, no spreadsheet required.
Send text or a link into Cortex from Safari, Notes or Messages. It waits in your journal for a quick look and can go straight into a thread. Widgets on the home and lock screen open a new entry in one tap, and Siri or the Action button can start a voice entry.
Face ID or Touch ID guards the app. Everything you have written exports as plain, readable files whenever you want them. Write-ups for a doctor can be produced in English, German or Arabic.
Cortex counts what you mention most and shows it to you, with nothing invented. Browse by subject, or by what you were doing when you wrote: observing, theorising, trying something, or holding on to a good moment. Release stays out of the browse view on purpose.
A plaque on the home screen shows one good moment from your own entries, a new one every three hours. Not a quote from someone else. Something you wrote on a better day, back when you need it.
Write that you will try something, and Cortex comes back later to ask. Did you get to it? Did it work, or not, or still testing? Your tap decides what the strategy becomes. The app never guesses.
Name a period: a trip, a new medication, a bad week. The moment opens as a portrait of that time, what you took, what came up, then the entries day by day. What was life like during that, as opposed to what do I know about it, which is a thread's job.
A third write-up is made for a model, not a person: a whole thread as one dated list, up to a hundred entries, with a short preface on how to read it. Paste it into any chat model and ask your own questions.
A demo journal with seventy-two entries over several weeks lets you try threads, Mimo and the write-up before you have written a line of your own. Switch it off and it is gone, your journal untouched.
Sleep and heart rate from the Health app sit beside the day as context, if you allow it. Cortex shows them. It never draws conclusions from them.
This is a real thread from one journal, seven months of it. Every entry in it is about teeth. But "coffee, then the pain" and "less pain after pantoprazole" are not the same kind of sentence, and a dentist should not have to read them as one pile.
So each entry gets a type at the moment it is written. Observation for what you noticed. Insight for what you suspect. Strategy for what you tried. Release for letting it out. Positive for what held.
The lines show where the 46 entries went. Read the dates across the cards and one story appears: a suspicion in March, a bad night, things tried in April, the coffee trigger noticed, and in May something that helped.
Each line is the one-line summary Cortex keeps for an entry, not the entry itself. Shown with the writer's permission, chosen to carry no names, places or people.
Every entry already carries its type, so a share is a filter, not a rewrite. Doctor: noticed, suspected, tried, helped. Yourself: everything. A friend: how it felt and what helped. Social media: only what helped and what you tried. Change the ticks and the sheet follows.
What the dentist gets:
Three visits in this thread went nowhere: one clinic found nothing, one visit happened unprepared and under overload, one took three hours for a generic answer. Each time the seven months of timing stayed in my head. The sheet puts it on paper before the appointment starts.
Tick Release and the hard days come back, for a friend or for me on a better day. One tap keeps them out of the dentist's copy. Cortex never says what caused what. It remembers and sorts. The reading is the reader's job, and yours.
Meet Mimo
Mimo reads everything you've written. It doesn't give generic advice, it answers from your own history and pushes back when you're jumping to conclusions.
Cortex uses large language model architecture as a conversational, low-demand interface. Rather than forcing structured forms, the system works from what you actually produce: a voice note, a short text, a line from the Watch.
The AI models run on AWS Bedrock in Frankfurt, through a small Cortex server that holds the AWS credentials so the app never does. AWS does not keep the text it processes. Voice transcription and search ranking run on your device; entry classification and chat are processed in the cloud.
Before any text is sent for processing, direct identifiers, your name, email, phone number, and postal address, are redacted on a best-effort basis. AI features require connectivity; there is no on-device language model.
Cortex for the Mac is coming soon, the same journal kept in step through your iCloud. Say something into the phone on the way home, and it is on the Mac when you sit down, already sorted and searchable. Write a long entry at the desk, and it is on the phone the next morning. Ask Mimo from whichever is in reach.
Raise your wrist and say what is happening. The words are turned into text on the watch itself and land in your journal a moment later. Today's medications are on the wrist too, and a dose logs with one tap. For the days when finding the phone and unlocking it is already too much.
Sync runs through your own private iCloud, tied to your Apple Account. I cannot see it, and there is no Cortex server holding your journal. What travels is your text, the sorting already done for it, threads, notes and intake logs. Voice recordings stay on the device that made them. One switch in Settings clears the iCloud copy again.
Explored how AI can support neurodivergent self-tracking. Validated the core concept: free-form journaling with automatic classification and pattern extraction. Defined the architecture and data model.
The iPhone app with on-device voice, the five entry types, Mimo, threads, the doctor write-up and daily intake. Over seven hundred commits, one user, every day, and still changing every week. Along the way a whole causal-inference layer was built, measured, and removed.
A Watch app for journal entries and medication doses without the phone, and the same journal on the Mac and iPad, kept in step through iCloud. The Mac version is coming soon. Built and running. Being tested on my own devices before anyone else sees them.
Public beta on TestFlight for early testers from the neurodivergent community. Gathering feedback and preparing for App Store submission.
Exploring Apple Intelligence and on-device foundation models for lightweight tasks that can run without the cloud, reducing how much text needs to leave the device. Not yet part of the app today.
About
I'm Mohamad Sakkal. I'm autistic with ADHD, and I've spent years in the slow grind of burnout, the kind where social connections fade because keeping them feels impossible, where the executive function to start small tasks just isn't there, where sensory overload turns a normal day into something you have to recover from for two more. The things that worked last month are gone by next month. I forget my own patterns. I forget what helped. I make the same mistakes I made a year ago because the memory of figuring them out the first time is gone.
I've always envisioned an app that does what my brain can't. Voice journaling that doesn't ask me to organize my thoughts before I have them, automatic classification so I can find things later without tagging anything, a chat assistant that actually remembers what I wrote a month ago, intake tracking that doesn't punish me when I'm not consistent. Every app I tried did one or two of these things, but never all of them, and never in a way that worked for an ND brain. So I built it.
I built Cortex solo over several hundred commits, mostly in the evenings and on weekends. Along the way I made, and corrected, the same mistake the heavily funded apps in this space are still making: I tried to have the AI organize my notes for me, infer causal patterns, label everything. It produced confident-sounding nonsense that quietly biased every interpretation. I deleted that entire architecture and rebuilt the app around a simpler principle: store the user's raw words, retrieve them when needed, and let the AI read fresh every time. The journal is the memory; the AI is just a reader.
Cortex isn't a journaling app. It's a cognitive prosthetic. External memory and external organization for brains like mine. I'm releasing it for other autistic and ND adults because the world doesn't need another productivity app, but it does need tools built by people who actually live this.
I'm looking for testers, for research, and for further cooperation.
Testers. Autistic and ADHD adults who journal, or want to, and will use Cortex for a few weeks and tell me what breaks and what helps.
Research. Clinicians and researchers who want to study what a journal like this does for people like me: how the five kinds hold up over months, what a doctor actually gets from a share, where the risks are.
Cooperation. Clinics, universities, communities or other builders who want to take some of this further together. Pilots, shared data agreements, a feature you need, an integration.
Write to admin@ms-dev.app.
Raw words in, stored on the device. A search index over them, not a summary of them. An assistant that answers from the original entries every time, cites them, and never asserts a cause. Built for ADHD and autistic adults, by one.
System Failures
Standard habit-tracking applications are built on a foundation of compliance, daily check-ins, streaks, and prompts. For individuals with Pathological Demand Avoidance, each of these micro-demands triggers an autonomic avoidance response, causing immediate and permanent abandonment of the tool.
When users are experiencing autistic burnout, the cognitive overhead required to open an application, navigate menus, and fill structured input forms is physiologically impossible. Existing tools require precisely the neurological resources that burnout depletes.
Current applications apply a fixed interaction model regardless of the user's daily neural capacity. A neurodivergent individual's functional capacity fluctuates significantly across hours and days. No commercially available tracker adapts its interface demands to this reality.
Mimo RAG Engine
Mimo queries your full journal history via hybrid search (vector similarity + FTS5), then interprets from raw context. It cross-references timestamps, physiological data, and prior entries to challenge premature conclusions.
On-device speech-to-text via WhisperKit. No cloud roundtrip for transcription. Neurodivergent speech patterns parsed locally, then structured by the extraction pipeline.
Every input classified: observation, venting, insight, strategy, positive. Each type triggers different extraction rules and contextual rendering.
Retrieval-Augmented Generation over full journal history. Hybrid search, vector similarity + FTS5 keyword, retrieves relevant entries. LLM interprets fresh from raw text every time.
A thread is a user-stated question with a curated set of entries. New entries are routed in by term match and by embedding similarity against the thread's description, above a fixed threshold, and the user can always overrule the routing.
A thread renders as a dated, month-by-month account built from the per-entry summaries already produced at analysis time. No extra model call. Observations and the writer's own theories are kept in separate sections, with sources marked.
Medications, supplements, accommodations tracked with time-block scheduling. Multi-dose support, validation status lifecycle. Mirrored to watchOS for tap-to-log.
Text or a URL from any app arrives through the share sheet, with article extraction and an optional thread pick. Home and lock-screen widgets for quick log. A voice entry starts from Siri or the Action button.
Biometric lock with inactivity timeout. iOS Data Protection at rest. Full structured export: entries with type, threads, intake, chat sessions, as JSON and readable text. Handover documents in English, German or Arabic.
Subjects are the terms that recur across entries, counted locally with no model call. Stances group entries by the type assigned at classification, so history can be read by mode: observing, theorising, acting, anchoring. Venting is excluded from browsing by design.
Cortex utilises advanced large language model (LLM) architecture to serve as a conversational, low-demand interface. Rather than forcing the user to interact with structured forms, the system extracts behavioural data seamlessly from natural, ambient input, voice fragments, short text, or passive signals.
The inference pipeline is built on AWS Bedrock (EU, Frankfurt, region by default, configurable in-app), providing enterprise-grade model access with zero data retention. Voice-to-text (WhisperKit) and the search index run entirely on-device; classification, chat, and the text embeddings that power search are served by Bedrock.
Direct identifiers are redacted from text on a best-effort basis before it leaves the device. AI features require connectivity, the app does not run a language model on-device.
Technical Architecture Diagram
To be replaced with final architecture illustration
Background
Solo developer based in Vienna. Autistic with ADHD. Years of burnout, executive dysfunction, sensory overload, fading social connections. The standard ND adult experience that no journaling or productivity app has ever held onto correctly.
Always envisioned a tool that combined the parts no single app gets right: sophisticated voice journaling, automatic cognitive-type classification, retrieval-based chat over your own history, intake state tracking, all local-first. Built Cortex because that combination didn't exist.
Original architecture used an AI-extracted causal knowledge graph as the context layer. After 33 days of escalating fixes, recognized that AI-inferred labels fed back into prompts produced manufactured certainty, not insight. Removed 126 commits worth of graph code. Rebuilt around RAG over raw text. Store the user's words, retrieve them on demand, let the model interpret fresh.
The design choices here come from being the user, not from market research.
iPhone, iPad and Mac run the same build and share one store through the user's private iCloud database. An entry captured on the phone arrives on the Mac with its analysis attached, so no device repeats work another already did. Voice audio stays local by design. On the Mac: Cmd+N for a new entry, Cmd+1 to 4 for the tabs. The Mac release is coming soon.
Record on the wrist, transcribe on the wrist, review, send. The text reaches the phone over the watch bridge and enters the normal pipeline. Today's intake list is mirrored to the watch with tap-to-log, and complications sit on the watch face.
Sync is Apple's CloudKit, tied to the Apple Account. No first-party server stores journal content. The only hosted component is a stateless AI relay that retains nothing beyond a usage count. Settings can purge the iCloud copy in one action; the local store is untouched.
Validated that unstructured journal input can be parsed into structured behavioural data. Defined extraction pipeline, classification taxonomy, and on-device storage architecture.
iOS app with cloud coprocessor (AWS Bedrock, EU). On-device voice, hybrid search, 5-type classification, RAG chat, threads, clinician handover, intake. 700+ commits, single user, daily. A causal knowledge graph and a Granger engine were built, measured and deleted.
Mac Catalyst target with CloudKit sync, watchOS app with on-watch transcription and intake logging. Built and running. Under test on the developer's own devices.
TestFlight beta for neurodivergent early testers. Gathering feedback before App Store submission.
Evaluating Apple Intelligence / on-device foundation models for lightweight inference, reducing data egress for tasks that don't require cloud models. Not yet integrated.
Two Perspectives
Natural shows how Cortex feels in everyday life, warm, intuitive, human.
Cortex maps your experience to how your brain actually works, through the language of neuroscience.