You open your AI when you need it.
You close it when you're done.
It forgets you in between.What if it never closed?

An AI that remembers who matters to you, tracks what you've committed to, and earns the right to act on your behalf, across every tool you already use.

For solo operators and prosumers running 3+ AI tools who keep re-explaining themselves every session.

Building an agent that acts in users' lives? There's a builders door →

The ceiling

Most people use AI the same way they use a search engine. They open it, ask something, get an answer, and close it. Today's models have started saving preferences and snippets, but that's a notepad, not a memory. There's still no real model of who's in your life, what you've committed to, or how you talk to your boss versus your kid.

A smaller group has gone further, connecting their calendar, their email, their tools. Written prompts. Scheduled them to run automatically. They're getting more out of AI than almost anyone.

But the cognitive work behind it, deciding what to automate, figuring out the right prompt, testing whether it works, maintaining it as things change, that's all still on you. And most people never get there at all.

The promise of AI was always that it would handle the repetitive, high-effort parts of work, so you could focus on the parts only you can do. That promise isn't fully here yet. Not because the AI isn't capable, but because it has no memory of you, no sense of who matters in your life, and no way to earn your trust over time.

Cortex is built to close that gap. Whether you're starting from scratch or you've already built sophisticated workflows, Cortex learns how you work, who matters to you, and what you want handled. It doesn't require you to be a power user to get power-user results.

A working theory of mind

Today's AI has mastered language. A mind needs more.

Language is just one part of a mind. A mind also remembers, learns from experience, understands relationships, makes judgments, and knows when to act versus when to wait. Most of those layers don't exist in commodity AI yet.

Language
Reading, writing, reasoning, conversing naturally.
From the model
Factual Memory
Remembering preferences you've stated and snippets you've shared. Flat, fragile, per-session.
From the model · emerging
Experiential Learning
Memory that evolves; rules that strengthen with use, decay when irrelevant, self-correct through feedback. Not a list of facts; a model that gets smarter over time.
Built into Cortex
Relationships
Understanding who the people in your life are, how they connect to each other, and how important each one is, resolved across every channel, updated continuously.
Built into Cortex
Judgment & Autonomy
Knowing what to act on, what to surface for review, what to ignore. Trust earned through demonstrated reliability, per person, per action, never granted all at once.
Built into Cortex
Accountability
A complete record of every decision, including what was done, why, and what to do differently next time. The ability to explain itself, and the honesty to pull back when it's wrong.
Built into Cortex

Built and operated by the founder. 90 days in production, n=1.

Other tools give AI a memory. Cortex is an attempt to give it a mind, one that doesn't just store what you tell it, but learns from experience, understands context, earns trust through its track record, and gets meaningfully better the longer it knows you.

A different kind of day

What this feels like at 6:30 AM, noon, and end of day.

Your brief is different today. Not because you rewrote the prompt, but because three things happened overnight that changed what matters. A key contact texted you last night, emailed this morning, and is on your calendar this afternoon. Cortex knows it's the same person. It knows the text was personal, the email is time-sensitive, and the meeting is high-stakes. It surfaces all three as one thread, ranked above everything else, without you asking.

  • People connected across email, messages, and calendar automatically
  • Relationships ranked by months of interaction history, not inbox order
  • Context from every channel combined into a single picture
  • Brief adapts to what's happening today, not a fixed schedule
Morning Brief, May 24
Urgent
Sarah Chen
Time-sensitive email re: Thursday deliverable. Conflict with 2pm prep meeting; she's also on your calendar at 4pm.
iMessageEmailCalendar
Review today
David Park
Replied to your Tuesday thread. Waiting 3 days; commitment window closing.
Email
Informational
Weekly digest
4 scheduling confirmations handled automatically. 0 corrections needed.

Between meetings, a single item surfaces: “You haven't responded to Marcus in four days. He's waiting on something you committed to last Tuesday.” You didn't set a reminder. You didn't write a rule. Cortex tracked the commitment from a message thread, noticed no reply came, and surfaced it at a moment when you're actually free to act, not mid-call, not at midnight. One tap drafts a reply in the tone you use with Marcus specifically, direct, brief, no filler. You approve it. It sends.

  • Commitments tracked automatically from messages and meetings
  • Surfaced at transitions, never during focus or meetings
  • Draft written in your voice, calibrated to that specific person
  • Say “from now on, always draft Marcus replies for approval”, one sentence becomes a permanent rule
Transition, 12:34 PM
!
Marcus Rivera is waiting on you
You committed to sending the Q2 breakdown last Tuesday. No reply sent. It's been 4 days.
“Hey Marcus, apologies for the delay. Attaching the Q2 breakdown now. Let me know if you need the segment view as well.”
Tone: direct, brief · Based on 23 prior messages with Marcus

A routine scheduling request comes in from someone you coordinate with regularly. The system has handled this exact type of request fourteen times without a correction. It just handled it. A request from someone more senior waits; the draft is ready, the context loaded, but it doesn't move until you approve. At end of day, you glance at the summary. One thing got slightly wrong. You correct it in one tap. The system updates the underlying rule, not just this instance. Tomorrow it won't make the same mistake. Not because you rewrote a prompt. Because it learned.

  • Trust graduated per person, per action; never a blanket setting
  • Higher-stakes relationships always surface for your review
  • Every action logged with full reasoning, always inspectable
  • One correction updates the root cause and propagates everywhere
Today's Actions, 3 handled
Reschedule, Jamie TorresHandled automatically
Moved 1:1 from Thursday 2pm to Friday 10am. Avoided your focus block and Jamie's conflict.
Trust: Level 3 · 14 of 14 successful
Reply draft, VP OperationsAwaiting your review
Draft reply to budget reallocation question. Tone: formal, thorough. Includes data from last month's report.
Trust: Level 2 · 3 of 5 successful
Why earned trust matters

The market is publicly proving “autonomy by default” doesn't work.

Three failure patterns from the past 18 months, identified by pattern rather than name, because the patterns are what matter.

Incident pattern

One agent platform shipped a CVE allowing 1-click remote code execution on every connected account. CVSS 8.8.

Incident pattern

An autonomous outreach agent sent 1,400 cold emails for zero replies; every recipient identified it as AI on first read.

Incident pattern

A subscription-based agent charged users $550 in overage fees they had no way to opt out of, with no path to cancel.

Cortex won't act until it's earned the right to, per person, per recipient, per action class. The agent-governance category that's now emerging scores the agent; Cortex scores the relationship. Trust is graduated, reversible, and logged. The longer Cortex runs, the higher the switching cost.

The moat
Founder telemetry · 90 days · n=1
≥90%
commitment-completion rate in the stated window
≥95%
trust-survival at Level 3 (no undo within 24h)
50%
drop in important-thread response latency by day 30

Directional, not benchmarked. Published transparently because we know what we're trying to prove, and want to be honest when we miss it.

How it works

One cognitive core. Your surfaces. Any AI client.

The same backend powers Cortex's own ambient surfaces and an MCP server callable by any LLM client. You don't have to choose.

Your surfaces
iOS appBriefings & approvals
Mac menu barSingle-item transitions
Web dashboardChat, memory inspection
Cortex cognitive layer
Memory · Rules
Trust · Ledger
Any AI client
ClaudeDesktop, mobile, Code
ChatGPTCustom GPT / MCP
Any MCP clientOpen protocol
Why now

MCP standardized this year. Any LLM client can call cognitive infrastructure through one protocol. Regulators are tightening on autonomous AI decisions; provenance, audit trails, and contest paths move from optional to required. The cognitive substrate category gets defined in the next 18 months by whoever has real time-under-load data; we have 90 days of it, and the corpus only emerges by running.

For builders shipping agents that act in users' lives

The cognitive substrate every consumer-life agent product re-implements, built once, MCP-callable.

We built Cortex because we were building a consumer assistant ourselves. Memory of people across messy life data, behavioral rules with lifecycle, per-pair earned trust, full action ledger with provenance; that's the scaffolding every product like ours has to ship before it can ship the actual product.

Consumer Cortex is the substrate's reference implementation. The same primitives become MCP-callable when the public server ships Q3 2026. Read Cortex's state and let your agent decide (state-only), or delegate to Cortex's policy engine and get back approve / human-required / reject (engine), per primitive.
Memory of People state-only
Query a unified people-graph across email, messages, and calendar.
Skip ~3 months of cross-source identity resolution.
Behavioral Rules both modes
Read the active rules and apply yourself, or call Cortex to evaluate them in context.
Rules with lifecycle, not prompt-string snippets.
Earned Trust both modes
Read per-counterparty (person, recipient, account) trust state, or have Cortex gate your agent's action against it. Agent-governance runtimes score the agent; Cortex scores the relationship.
Per-(counterparty × action class) graduation logic.
Action Ledger state-only
Append-only log with reasoning, cost, and reversal path.
Full provenance, GDPR Article 22-ready.
19 MCP toolsPublic server · Q3 2026Pull (action.pending) or push (Claude Code Routine → Slack/email/webhook)

After six months, it knows your rhythms.
Your relationships. Your priorities.
What you trust it with and what you don't.

Starting over means starting from zero.

This is being built now.

Founder-led. In active development. Public MVP July 20, 2026.

Get early access →