I designed an AI second brain companion that connects what no single app can see — looking for builders

#ai #opensource #privacy #sideprojects

Every personal tool you use is a silo.

MyFitnessPal knows what you eat. Fitbit knows your steps. YNAB knows your spending. Your journal app knows how you feel. Apple Health aggregates some health data. None of them talk to each other.

I kept running into this: I'd notice a pattern in my life — spending more on stressful weeks, sleeping worse after certain kinds of conversations, feeling sharper on days I ate clean — but no tool could confirm it. The data existed across five different apps. The connection existed only in my head. And my head is exactly the thing that can't be trusted to see its own patterns clearly.

So I designed Cortex.

What it is

Cortex is an AI companion that absorbs any signal from your life — health data, finances, mood, diet, social interactions, calendar, habits, whatever you feed it — and finds the cross-domain patterns you can't see from inside your own experience.

Not a health tracker. Not a habit app. Not a finance tool. A companion that sees all of it and tells you what it sees honestly.

The examples that made it click for me

When I was writing the spec, these are the moments where Cortex went from abstract to real:

The social one:

"You've logged a headache after 5 of 6 calls with person X. On weeks without those calls, you report zero headaches. I don't know what's happening on those calls, but your body does."

No single app would ever surface this. Your health tracker doesn't know who you talked to. Your calendar doesn't know you got a headache.

The financial one:

"You've logged impulse purchases on 6 of the last 8 days you reported high stress. Average spend on high-stress days: $84. Low-stress days: $22. This isn't about willpower. The stress is the trigger."

Your budget app shows you the damage. Cortex shows you the trigger.

The confrontation:

"You've been on screens past 1am four nights this week. Sleep averaged 4.9 hours those nights. Energy self-reports were all 1 or 2. You mentioned low motivation yesterday. This isn't a mystery."

Every wellness app would hedge: "Have you considered that your screen time might..." Cortex doesn't do that. When the data is clear, it says it clearly.

Key design decisions

Runs on local models. Gemma, Llama, Mistral — whatever fits your hardware. Your most intimate personal data (mood, finances, health, relationships) never leaves your machine. Privacy is structural, not policy.

Input is multimodal. Voice, text, photos. Snap your lunch — that's a log. Screenshot a bank notification — input. The friction should be as close to zero as the thought itself.

MCP for integrations. Any data source with an MCP server becomes an automatic signal feed. Bank feeds, Google Calendar, fitness apps. The community builds and shares connectors.

Store everything. Act on patterns. Never interpret meaning. This is the core principle. Cortex will say "you've logged feeling low 4 of 5 days and your spending spiked the same week." It will never say "you have a shopping addiction." The line is observation vs. interpretation. Cortex holds the mirror. You bring the meaning.

Entity resolution. You log "bodyweight stuff" one day, "bodyweight workout" the next, "floor exercises" a week later. Three names, same activity. Most systems fragment this silently and you never know. Cortex stages unknown inputs, clusters them nightly by semantic similarity, and resolves them — either automatically or by asking you once.

Brain lens modes. The brain has known bugs. Each lens patches one:

Antibrain — patches guilt spirals. Shows your return rate, not your gap length.
Beautybrain — patches progress blindness. Shows the distance from where you started.
Flowbrain — goes quiet when analysis would be interference. Still records. Speaks when asked.

What this is right now

An idea. Fully specified in one self-contained document — architecture, behavioral principles, entity resolution, correlation engine, interaction examples. No code yet.

I'm looking for:

People who want to build it — pick a component and start
People who want to extend the idea — new lens modes, domain-specific lenses (finance, athletics, chronic illness), MCP server ideas
People who want to challenge it — if something is naive or won't survive reality, say so
People who want to propose architecture — better approaches to the correlation engine, graph storage, fingerprint generation