The memory layer for the agentic workforce

A brain for your company that never forgets — and never quits.

Your agents and people do the work. Cortex remembers all of it — so you can recall anything in milliseconds, and talk back when a milestone needs you.

Get started — it's one command Watch the demo
Devs & solosTeams & agenciesTrading desksGame studiosSRE / opsStartups shipping agents
cortex — demo
npx create-cortex trading-bot ✔ local CockroachDB running · 127.0.0.1:26257 cortex save "bought DOGE on momentum — avg 0.22, stop -2%" ✔ remembered · embedded · 0.18s cortex ask "why did we buy DOGE?" → "bought on momentum — avg 0.22, stop -2%" (similarity 0.94 · 3 months ago) cortex ask --approval "increase position? (risk +0.5%)" 📱 pinged 1 phone · waiting on your call…
3
alert types only
100%
recorded & auditable
survives a node outage
<50ms
semantic recall
The demo · 3 min

Watch memory hold the line while production doesn't.

A trading platform, a dev team, and a game where the creature remembers you — one shared brain. We prove it: kill the database node, and the memory — and the people relying on it — don't even notice.

Demo video — recording in progress

Capture → recall → agent approval → kill-the-node resilience. We're producing the 3-minute walkthrough now.

Placeholder · to be replaced
Get started in seconds

One install everyone knows. Then it just remembers.

The front door is a command you already know — npm. No database setup, no cloud account, no signup. And it plugs into the CLIs you already live in.

🧱

Install

The universal one-liner. Sets up Cortex + a local CockroachDB, zero accounts.

$ npm install -g cortex
🔌

Connect your CLIs

One MCP server covers them all — Claude, Cursor, Codex, opencode — no per-app plugins.

$ cortex connect --claude
🔍

Ask

Semantic recall of everything it captured — in milliseconds.

$ cortex ask "how did I fix auth?"
📲

Act

Agents do the safe work; you approve the milestones from your phone.

$ cortex ask --approval "deploy?"
Works inside the CLIs you already use
Claude CodeAntigravity · agyopencodeKimi K3QwenCodexCursor (MCP)

$ cortex connect <cli>  ·  one MCP server behind it (capture · recall · ask · approve · note) — no per-app plugins

Not a log. A living brain.

See your company brain — every memory, mapped.

Context windows are temporary. RAG helps an agent answer better; memory helps it behave smarter. Cortex visualizes that memory as a graph the whole company builds.

PublicPrivate
shared memory yours only
🌐

Public — grows the company index

Contribute to the shared brain. Visible to teammates and agents, searchable by everyone — a crowdsourced datastore of what your company knows.

🔒

Private — yours, and kept

Contribute quietly. Retained for you, never added to the company index. Your footprint stays yours.

🔗

Relations, not rows

Memories connect to memories. The graph shows what depends on what, and who knows it — so nothing important goes quiet.

🧠

Built on Three.js

An interactive constellation you can fly through — click a node to open the memory behind it. Visuals like this are why the demo feels alive.

The value

We bridge the gap between what agents are doing and what your people know.

Agents produce a flood of events. People can't absorb a flood — so the ones who matter never know what's happening. We close that gap with signal, not spam.

🤖

What your agents are doing

A constant stream of micro-events: tools called, files changed, steps attempted, context switched.

tool · deploy_184 · build ✓tool · migrator · 0.32ssub-step · retry (2)context · switched → api
THE
BRIDGE
signal, not spam
🧍

What the human needs to know

Only the moments that cross a decision — started, needs you, finished differently.

→ START · deploy 184➜ NEEDS YOU · approve?✓ DONE · shipped · 2m
The philosophy

Milestones, not noise.

We record everything — but we only alert you on the few moments where being told changes your outcome. Two regimes, one system.

Silent
Agent ran a safe step in the background — build, retry, query, a happy-path sub-task.
Recorded · never a ping
Alert
Crossed a milestone — a phase completed, or "needs your decision".
A push you can act on
Priority
Finished differently than expected — failure, delay, or an answer you'd want tonight.
Wake you if needed
It's a conversation, not a notice

Your phone is a control surface for your agents. Not a pager.

The push is the doorbell. The thread is the room. You don't just get told — you ask back, redirect, and decide. The whole exchange becomes memory the agent recalls next time.

🤖 Deploy Agentjust now
Deploy commit 8e7fc2a to production?
run #184 · risk: low · smoke tests pending
ApproveWait
You · asking back
"What does it need me for?"
It explains: "Smoke tests passed. 14 msgs behind. Rolling out; I'll verify & report."
Reply — ask anything, or approve…
✅ Deploy Agent+4m
Shipped · 2m 11s · smoke green. Saved to memory.
1

Doorbell

An agent hits a milestone and pings you — "needs your decision", not a firehose.

2

Ask back

Tap through to the thread. "What does it need me for?", "what if I say no?" — it answers with real context.

3

Decide

Approve, edit, or deny from the phone. The decision is stored transactionally.

4

It resumes — and remembers

The agent executes the safe work and the whole exchange becomes memory it recalls next time.

One memory, any product

Same engine, four wildly different businesses.

The memory layer is universal — proven by live integrations on the same CockroachDB core: a marketing agency, a dev team, a trading platform on a Raspberry Pi, and a game where the creature remembers you.

📣 Marketing agency🧑‍💻 Coding sessions💱 Trading🎮 Ocean Dreams
Use case · live

A marketing agency that never forgets a client

Every account person and their AI captures client context — preferences, brand voice, campaign history, reporting cadence — into a shared per-client brain. Onboarding a new account lead takes days, not months, and brand voice never drifts when someone leaves.

  • Client memory — "what tone does this client hate?" answered instantly.
  • Onboarding — new team member asks the brain, not the person who knows.
  • No drift — brand guide and voice persist even as people leave.
  • Shared — one brain per client, private per-contributor on request.
Try it
$ cortex save "client prefers short, no-jargon copy"
$ cortex ask "what did Acme reject last quarter?"
$ cortex note "--client acme" "avoid 22:00 sends"
The product bible

What we believe. What we're building. What agents must stay true to.

This page is the living contract: humans read the story, agents read the spec below. Ten immutable principles keep everyone — human and agent — on the same page while we build.

01

Memory is the product

Not an afterthought bolted onto an app — it is the reason the product exists.

02

Memory must never go down

An agent whose memory goes offline doesn't degrade gracefully — it stops.

03

Bridge, don't flood

We close the gap between what agents do and what people know — with signal, not spam.

04

Milestones, not noise

Record everything. Alert only on start, needs-you, and done-differently.

05

Human decides, agent executes

Autonomous on safe work; consequential decisions stop for a person.

06

One core, any product

Swiss Army Knife. The memory layer is universal — trading, games, apps, code.

07

Brain health, not timecard

Participation is vitality, not surveillance. We surface who's feeding the brain.

08

The front door matters

One command, zero accounts to see the magic. Onboarding is the demo.

09

Memory improves itself

Consolidation makes it measurably better over time — a learning curve, visible.

10

Kill the node, memory survives

Resilience is recorded on camera, not claimed in a README.

The machine-readable spec is embedded in this page as JSON-LD (#cortex-bible) — agents can read it to stay on track.
Multi-tenant & brain health

One shared brain. Know it's actually being used.

Everyone logs into the same brain — and a healthy brain is one that's used. We surface who and what is actively contributing, so nothing important goes silent.

🔐

Multi-tenant & RBAC

Org namespaces, per-user audit, roles. Who did what — revocable, attributable.

🫀

Brain health

A pulse dashboard: people, agents, and systems actively feeding memory — so nothing important goes quiet.

Brain health · Eng org▼ LIVE
Agents96%
Engineers82%
Design44%
CI / on-call100%
Legacyoff
Pricing

Start free. Scale when your brain does.

Solo
$0
  • Local CockroachDB dev mode
  • Unlimited capture & recall
  • All 3 alert types
  • CLI + web inbox
Start free
Enterprise
Custom
  • SLA + multi-region GLOBAL
  • Self-hosted / VPC
  • Advanced RBAC & compliance
  • White-glove onboarding
Talk to us
Get started

The future of work is agents. Someone has to be in control — and know what they did.

CockroachDBAWSBuild with Agentic Memoryopen source
Build log · living

Progress, tracked here so agents and humans stay aligned.

This section is updated as we build — the page is the bible, the log is the current verse.

In build

Landing page · premium redesign

Refined dark theme, editorial type, live neural-network field, interactive memory map, full mobile pass.

Aug 3
In build

Product bible (10 principles) + machine-readable spec

Embedded JSON-LD so agents building the app can read the immutable truths and stay on track.

Aug 3
In build

Fresh MCP server (universal adapter)

Baked into the spec: one Cortex MCP server (capture · recall · ask · approve · note) covers Claude, Cursor, Codex, opencode, Copilot — no per-app plugins.

Aug 3
Next

create-cortex scaffold + CLI

npm install -g cortex → local CockroachDB + save / ask / ask --approval surface.

Next

Capture → recall end-to-end

10-second "oh cool" moment on local CockroachDB, then the hook video.

Next

Trading bot on Raspberry Pi + Tailscale

Wire the Solana platform on real hardware — the Swiss Army Knife proof.

Later

Ocean Dreams integration

The creature remembers player behavior across sessions — a demo that visibly "gets" your memory.

Later

Agents layer + resilience

Operator agent, approvals, CDC → SNS → push, kill-the-node.