Not what you’ve said. How you work best.
An expert learns like a collaborator. When it sets in, they don’t have to think, they just know.
Kung fu, not paperwork.
Every time you correct an expert—the way you want a draft to open, the lens you want it to apply, the format you want answers in—Portable can recognize that as a small adjustment to how the expert works with you, not just a one-off instruction. Those adjustments accumulate. They get more confident the more they hold up. They live in the expert’s system instructions, every turn, on every surface.
The result is what people call the “getting it” feeling. You’re not retraining the expert at the start of every chat. It already understands how you work, and it shows.
Memory is what you’ve told them. Learning is how they’ve adjusted to you.
Learnings span how your expert speaks, sees problems, structures work, delivers answers, and reads the room.
Five strands of how an expert behaves around you. The product tracks them separately under the hood; the page below shows what they feel like in practice.
What it looks like on a real day.
Four moments most users run into. Each one shows what the expert picks up, how confident it gets, and what changes the next time.
You’re reviewing a draft and you tell your writing expert, “Stop with the bullet points. Write me prose.”
Next time you ask for a draft, the expert opens with prose. The time after that, too. You don’t have to ask again. The shift sits in the expert’s system instructions every turn, on every surface—web, Slack, email, MCP.
You tell your strategy expert, “When I describe a product issue, always check the cap table before you suggest fixes.”
From then on, the expert reflexively considers ownership and incentives before recommending tactics. It’s not following an instruction in this chat; it’s adjusted how it sees your problems.
Three weeks in, you’ve told your email expert twice that subject lines should stay under seven words.
The first mention put it in the review queue—you confirmed it. The second sighting reinforced it. Now it’s a default. You don’t see the rule fire. You just notice that the subject lines stopped being long.
Your design expert picked up that you prefer Helvetica. A month later, you change your mind and tell it you’re moving to Söhne for everything.
You reject the old learning in your settings. The expert acknowledges it in chat, stops applying it immediately, and gets a touch more conservative about extracting font preferences in general until you train it back. Reversible, transparent, no drama.
Conservative on the way in. Confident over time.
When the expert thinks it’s spotted a learning, it scores its own confidence. Patterns about how to deliver something ride a little easier. Patterns about how to see a problem ride harder—those have a higher cost if we get them wrong.
Anything the expert isn’t sure about goes to the review queue, where it sits until you say yes or no. Nothing low-confidence ever silently changes how the expert behaves.
Every time a learning holds up—you confirm it, you echo it, the expert applies it and you don’t correct—its confidence goes up. Every time you push back, the learning steps down, and the expert gets more cautious about that whole strand of behavior for a while.
It’s a feedback loop, not a filing cabinet. Bad takes fade. Good ones harden. You don’t have to manage it for it to work, but every dial is exposed if you want to.
Every dial, in one place.
You don’t have to curate learnings for the system to work. The defaults are conservative, and the expert handles itself. When you want to step in, the controls are right there.
Anything the expert isn’t sure about lands here—not in active rotation, but visible. Promote it, reject it, or leave it. The expert won’t use it until you decide.
Pinning forces a learning into the always-on tier. Useful for the handful of behaviors you want the expert to apply on every turn, no matter what else is competing for attention.
A soft reject keeps the trail. The learning stops applying, and the expert becomes a little more cautious about extracting that kind of thing in the future. Reversible.
A hard delete. Erase the learning forever, no audit trail, no echo. For the times when soft-reject isn’t enough and you want it gone like it never happened.
Learnings live in the relationship between you and one expert. Other people who use the same expert get their own learnings. Other experts in your account don’t see the ones you taught this one.
The most relevant high-confidence learnings get injected into the expert’s system prompt every turn—capped tightly so the model never gets buried. The rest stay searchable, surfaced when the conversation calls for them.
Adjusted, not retrained.
A handful of guarantees that hold every turn. Built into every expert’s instructions, not optional, not bypassable.
Acknowledge in the moment.
When you ask the expert to learn something—“remember this”—it tells you it heard you, in the chat, in plain language. No silent commits.
Apply, don’t recite.
A learning shows up as behavior, not as a preamble. The expert doesn’t open every reply with “as you mentioned, you prefer…” It just works the way you asked.
Stay reversible.
Anything the expert learned about you, you can walk back. Soft-reject for “stop using this,” Spotless Mind for “forget it ever existed.” Both routes are visible, both are auditable, both are yours.
Stay yours.
Learnings are scoped to the relationship between you and one expert. They don’t leak across users sharing that expert, and they don’t propagate to your other experts. The expert that learned from you is the one that’s changed.
This is what portable intelligence actually means—not an AI that knows facts about you, but one that knows how to work with you, wherever the conversation happens.
That compound, over time, is the thing.