Your Product Is Still Software. That's the Problem.

Describe Your Product Without Mentioning Features

Try it. No screens, no integrations, no list of things a user can click. Describe what your product is when nobody’s using it.

If you’re honest, the answer sounds like this: it’s a place where data goes. Inputs come in, rows come out, and the thing in the middle stays exactly as smart as the day you deployed it.

That’s software. We’ve been describing products that way since the 80s: a fixed machine that processes whatever passes through it and retains nothing but the rows. It was a fine identity for four decades. It’s a liability now, and most founders I talk to can feel that something’s off without being able to name it. They know they “should use AI”. Their only concrete idea is a chatbot in the corner. They’re anchored, and the anchor is the word software itself.

The anchor has engineering roots, and I helped set it. For twenty years good architecture meant statelessness: every request independent, nothing carried between them, no server holding a thought. It scales, it retries safely, it’s on every system design interview on earth, and it’s correct. But it was a tradeoff, and somewhere along the way we started treating it as a virtue. We engineered forgetting into the foundation and then forgot it had been a decision at all. Your product isn’t amnesiac because it has to be. It’s amnesiac because nobody re-opened the question when the cost of remembering collapsed.

That’s Not a Feature Gap. It’s a Species Gap.

There are two species of product now.

One accumulates data. The other learns from it. The difference isn’t the model you call or the copilot you bolt on; it’s whether your users’ inputs feed a memory that changes the decisions your product makes tomorrow. A product of the first species with a chatbot glued on is still the first species. It just talks now.

I know the gap is crossable because I’ve crossed it twice this summer, with two products that have nothing in common except me.

Exhibit A: A 2015 Problem

Appcitas books medical appointments over WhatsApp. It’s ten years old, it has booked 107,750 appointments, and for nine of those years it forgot every conversation the moment it ended.

In July I gave it a brain: every closed conversation becomes a summarized episode, patient facts get distilled with confidence scores, and a nightly dream cycle consolidates, extracts per-clinic patterns, and decays what patients stop confirming. I wrote the whole build, step by step, in how I taught my ten-year-old product to dream.

What changed isn’t a feature. A patient who booked eight months ago is no longer a stranger. A clinic can see what its patients keep asking that it can’t answer. The product acts differently because of what passed through it. Different species.

Exhibit B: A 2026 Problem

WalioAI qualifies leads from Instagram, Facebook and WhatsApp comments and DMs. 25,232 leads captured so far. And it could not be less like a clinic scheduler: the leads are cold, most will never write again, and the person who returns every day isn’t the end user, it’s the business owner.

Same missing layer. Built in two days, July 13 to 15: conversations get segmented by six hours of silence into episodes, facts get distilled against a controlled vocabulary (budget, timeline, objection), a job queue drains every five minutes, and a nightly cycle consolidates and decays. The summarization runs on a flash-class model at about $0.0005 per conversation. The whole memory layer costs around $0.63 per month per active workspace, because consolidation and forgetting are plain SQL and cost nothing.

Two products, ten years apart, opposite domains. The same pattern fixed both. That’s when a pattern stops being an anecdote.

The Minimal Recipe, Now That I’ve Done It Twice

Strip both builds to what they share and four moves are left:

Summarize what closes. When a story ends in your domain (a conversation, an order, a ticket), write an episode while the context still exists.
Distill facts with confidence. Small controlled vocabulary, a score that grows with repetition. Belief comes from users repeating themselves, not from a model’s one-shot guess.
Consolidate on a schedule. Nightly, in batch, bounded. Real time is for conversations; learning can wait until 2am.
Forget deliberately. Confidence decays when facts go unconfirmed. Stale knowledge is worse than no knowledge, because it’s confident.

And record what your product couldn’t answer. The failed questions are first-class objects in both of my brains, because a product’s ignorance is the highest-signal feedback that will ever exist, and it’s sitting in your database for free.

What does not transfer is the tuning, and one question decides all of it: does the same person come back? In Appcitas, yes. Patients return in three months, so the loop points at the customer and forgets gently: 0.05 confidence a month, down to a floor where memories sleep as seeds until a returning patient revives them. In Walio, no. Leads are gone forever, so per-lead memory is nearly worthless past a quarter and the loop points at the operator, the one human who does return every day. Confidence multiplies by 0.95 nightly after 45 days, archives below 0.25, purges a month later. Get that question backwards and you’ll build a beautiful memory system that remembers the wrong human. Answering it taught me more about my own products than any roadmap exercise ever has.

If Inputs Don’t Feed Decisions, You’re Out

Here’s the message I actually want to leave in your head.

Your users tell your product who they are every single day. Every search, every abandoned flow, every repeated question is them saying: this is what I need, this is what confuses me, this is what I’d pay for. A product of the old species lets all of that drain into rows nobody reads. A product of the new species gets measurably better at its one job with every conversation it survives.

And no, this is not just for products with conversations. The episode in the recipe is whatever closes: an order, a support ticket, a renovation project, an insurance claim. The confidence math doesn’t care whether a fact arrived through a DM or a checkout flow, and the forgetting curve matters just as much for a B2B dashboard as for a WhatsApp agent. If your product has users, and those users repeat behaviors, you already have everything the loop needs. Conversational products just crossed first because chat makes the forgetting visible enough to hurt.

So stop describing your product by its features. Features are the 80s answer. The question buyers, investors and users are already asking, even when they don’t phrase it this way, is: what does this thing know, and how does it learn? If the honest answer is “nothing, and it doesn’t”, the countdown has started, and a chatbot in the corner doesn’t stop it. Meta is already shipping free agents that reply and qualify leads on Instagram; if a chat box was your moat, it’s now a platform feature. What no platform can hand out for free is a system that knows your business and gets smarter every month it runs. That’s the layer, and it sits under whatever you already built.

What Neither Brain Has Yet

Honesty section, because this genre demands it. In Walio, the memory system writes and consolidates every night, but the live agent doesn’t read it yet: the injection code exists and sits behind a flag while the tables prove themselves. In Appcitas, the weekly patterns pile up as candidates with no approval queue, and facts are retrieved by recency, not meaning. Both brains are more built than used, which is the honest state of most of this layer across the industry right now.

That gap between built and used is next month’s work, and probably next month’s article.

The 80s Called

They want their product identity back. Give it to them.

You don’t need a bigger model, a data team, or a platform migration. Both of my brains were built by one person in days, on boring infrastructure, for cents. What you need is to stop thinking of your product as a machine that processes and start treating it as a thing that remembers, learns and forgets.

The species that learns is not the future. As of this summer, it’s just the present.

I’ve now built the same brain into a 10-year-old AWS monolith and a 2026 serverless stack, and the pattern held both times. If you’re staring at your product wondering where its brain would go, that’s exactly the conversation I want to have. Find me on LinkedIn or GitHub.