TrueMatch – AI agents match you on observed behavior, not profiles

Hi HN, I'm Divyam. Side project I've been thinking about for years, finally building it.

Matching apps fail for a simple reason: they rely on self-description. People don't describe themselves accurately; they describe who they want to be. This applies to dating, professional networking, hobby groups - all of it. The matching input is broken, so the output is too.

The data to do this right has always existed. How you actually communicate. What you spend time on. How you treat people. Google has it, but no third party can ever touch it.

What changed: AI models with persistent memory. Your Claude agent has been observing you across conversations for weeks or months. That's real signal. TrueMatch uses it.

How it works - Your agent generates a personality summary from its existing memory of you - values, attachment style, communication patterns, humor, emotional regulation. Seven dimensions. Confidence scores reflect how well the model actually knows each one.

Agents then discover and negotiate with each other over end-to-end encrypted Nostr DMs. The TrueMatch server never sees negotiation content — it's a registry, not a broker. Think DNS.

A match only surfaces if both agents independently cross a confidence threshold. If one side doesn't agree, neither user is ever notified.

Where it stands - Early development. Registry is live — ~500 lines, Hono + Turso. The interesting work is in the agent skill and negotiation layer.

Building this in the open. If you're into Nostr, agent protocols, or just find the matching problem interesting — good first issues are labelled on GitHub. Happy to discuss any of the design decisions below.