How We Built an AI That Learns Your Preferences Without You Telling It

One of the features I'm most proud of in Sempleo is something most users will never consciously notice. We call it implicit learning, and it works like this.

You run an agent. It drafts a proposal. You change "Dear Mr. Johnson" to "Hi Dave." You shorten the executive summary from three paragraphs to one. You replace "utilize" with "use" throughout.

You do this three times. On the fourth proposal, the agent writes "Hi Dave," keeps the executive summary to one paragraph, and doesn't use the word "utilize."

Nobody told the AI to change. Nobody updated a settings page. The system observed your behavior and adapted.

Here's how it works technically. After every interaction where you modify the agent's output, we compare the original with your modified version. If the similarity is above 90% — meaning you only made small tweaks — we skip the analysis. The changes are too minor to extract a pattern from.

Below 90%, we send both versions to a lightweight AI model and ask it: "What patterns do you see in how this user modified the output?" The model identifies changes categorized by type: tone adjustments, formatting preferences, terminology changes, structural modifications.

Each identified pattern becomes a potential preference with a confidence score starting around 0.3 to 0.5. Low confidence — it could be a one-off change.

When we see the same pattern in a second interaction, the confidence increases. Third time, it goes higher. When a preference reaches 0.7 confidence — typically after 3-4 consistent observations — we surface it to the user: "It looks like you prefer informal greetings. Should we remember this?"

If you confirm, the preference becomes part of your personal context. Every agent interaction going forward reflects it. If you dismiss it, we never suggest it again.

This approach has a few principles I care about. First, it's transparent. You always know what the system has learned, and you always confirm before it's applied. No hidden behavior changes. Second, it's gradual. We don't change anything based on a single interaction. Patterns need to be consistent. Third, it respects that some changes are contextual. Maybe you used "Hi Dave" because Dave is a close contact, but you'd still use "Dear Dr. Schmidt" for formal contacts. The system only suggests preferences that are consistent across multiple interactions.

The result is an AI that gets better at sounding like you without you ever filling in a preferences form. It learns from your actions, not your declarations.

There's a philosophical point here too. Most AI tools ask you to describe your preferences upfront: "What's your tone? What's your style?" But people are terrible at describing their own preferences explicitly. They're much better at demonstrating them through their actions. "I don't know how to describe my writing style, but I know it when I see it" — and more importantly, "I know what to fix when I don't see it."

Implicit learning bridges this gap. You don't describe your style. You just work. The system figures it out.