One Chrome extension captures every AI conversation. One MCP server feeds decisions, preferences, and context back to your IDE. Your context stops dying between sessions.
Free during beta · No credit card · Local-first & private
How it works
Install the extension. Sign in once. Conxt captures every AI conversation in the background — and your IDE sees the full history.
Add Conxt to Chrome. It works immediately on Claude, ChatGPT, and Gemini — no configuration.
Use your AI tools exactly as you do today. Conxt listens silently and extracts structured memory — decisions, preferences, entities — from every session.
Connect Cursor or Windsurf to your MCP endpoint once. From that point every new session starts with your decisions, preferences, and context already loaded — no button, no copy-paste.
IDE Integration
Conxt exposes your memory graph as a remote MCP server. Point Cursor or Windsurf at one URL — your decisions, preferences, and context load automatically before every prompt.
https://mcp.conxt.dev/mcp/LiveOne permanent API key works across every IDE. Generate it once at conxt.dev/connect-ide.
For teams
Every decision your team makes — in Slack, in Claude, in Cursor — captured once and available to everyone. No more “wait, what did we agree on?”
Features
Built for developers who live in multiple AI tools every day and are tired of re-explaining themselves.
7 memory types captured automatically
Pricing
No surprise charges. Lock in beta pricing when you upgrade.
Free
$0
forever
For individuals getting started with AI memory.
Pro
$9
per month — coming soon
For power users who live across multiple AI tools.
Team
$25
per seat / month — coming soon
For teams that want shared context and memory governance.
Founder
Founder, KPZG LLC — Kalamazoo, MI
I built Conxt because I kept re-explaining myself to AI tools that should already know me. Every time I switched from Claude to ChatGPT to Gemini, I started from scratch — same stack, same preferences, same decisions, re-typed every session.
So I started capturing my own sessions to measure how bad it actually was. What I found was worse than expected: 4,189 memory records in my own graph, and over 60% were duplicates — 604 separate records describing the same product, written 585 different ways. Not identical enough for a hash to catch. Paraphrastic duplication that only semantic matching can resolve. I wrote a research paper about it.
I work in telecom retail during the day and build on nights and weekends. I'm finishing an MS in Computer Information Systems at Colorado State. Conxt is the infrastructure I wanted to exist — built solo, in public, measured on real data.