How agent memory compounds.

Every AI tool demos beautifully on day one. The real question is what it's like in week six — when the honeymoon is over and the value has to come from somewhere other than novelty. For most tools, week six feels exactly like day one, and that's the problem. You're still explaining your business from scratch, every session, forever.

We call this the amnesia tax. It's invisible on a feature list and enormous in practice: every re-explained policy, every re-discovered pattern, every insight that evaporated when the chat window closed. A tool that forgets you can only ever be as good as its model. A system that remembers you gets better than its model — on your store specifically.

The durable advantage isn't the model. It's the accumulated, structured knowledge about your business.

Lumnya's memory isn't one big transcript. It's three deliberately different layers, each with its own rules:

• Agent notebooks — each specialist keeps a private record of what it has observed. It can share entries with the team, but no other agent can overwrite its notes.
• The shared book of facts — when an observation is verified, it's promoted to a team-wide fact, stamped with which agent found it and how confident it was.
• The fresh snapshot — agents read a current, private copy of your store data instead of hammering your live systems, so they're fast and gentle on your Shopify limits.

Here's the loop in miniature. The Inventory Forecaster notices hoodies selling 2.1× faster once temperatures drop — an observation in its notebook. After enough evidence, it becomes a shared fact: “hoodie velocity doubles in winter,” author: Inventory Forecaster, confidence: 0.92. Months later, a pricing recommendation cites that fact instead of re-deriving it — and when you approve the reorder it justified, the outcome strengthens it. One discovery, written once, working everywhere.

Multiply that by every agent, every scan, every approval you grant or refuse, and you get the compounding curve: each cycle starts from a richer picture than the last. Week six doesn't feel like day one. It feels like working with someone who's been on your team for six weeks — because, in the way that matters, it has.

A memory this valuable raises an obvious question: whose is it? Ours is the boring, load-bearing answer: yours. Everything an agent learns is isolated to your store — never pooled across merchants, never used to advise a competitor, exportable and deletable on request. The moat this builds belongs to the store it's about.

What we're working on next in the memory system, honestly labeled as in-progress: confidence decay (facts should age unless re-verified), contradiction detection (two agents disagreeing should surface, not silently coexist), and a memory inspector so you can browse what your team knows and correct it directly.

That last one matters most to us. A memory you can read, question, and edit is the difference between intelligence that belongs to you and intelligence that merely lives near you. If you want to help us pressure-test it, the private beta is exactly for that.