How Twinly remembers you

Under everything sits an always-on memory. The twin continuously learns who you are, not only when you tell it to remember something. Two signals feed it, both captured automatically:

• Your corrections — the strongest signal. A steer like never text after 11 or I prefer the aisle seat becomes a standing rule the twin follows from then on.
• What your completed tasks reveal — after a task finishes, the twin pulls a small number of durable facts: a stated preference, a named relationship, a vendor you use.

It is deduped, so restating something it already knows does not pile up. It is fed into every task, so the twin applies what it knows about you no matter which model is running.

Memory is append-only and backed up: a bad write can never wipe what the twin knows about you. It accumulates a richer picture of you over time, the way a real assistant who has worked with you for years would. The only thing that clears it is an explicit reset by you.

Separately from facts about you, the twin remembers how to do tasks it has done before. The first time you ask for something new, it figures it out; after that it replays the steps it learned, faster and at no AI cost. This is the Pathways layer described in How Twinly executes.

Learned pathways are self-healing: if a site changes and a saved path stops working, the twin notices, drops that path back to unproven, and re-learns a corrected version on the next run. You do not have to maintain anything.

The twin also looks ahead, locally and quietly, and offers help before you ask. These suggestions are read from your own calendar, mail, and saved notes, never by running an AI in the background:

• A meeting soon → it offers a quick prep brief.
• Two events overlapping today → it offers to sort the conflict.
• Something you said you were waiting on, now overdue → it offers to draft a polite chase.
• A deadline coming up, or a spike of unread mail → it offers a reminder or a catch-up.

When a task fails, the twin remembers why, what it tried, where it got stuck, what the obstacle was (a login wall, a canvas it could not target). The next time it faces the same situation it adapts instead of repeating the same dead end. Over time the twin not only learns what works for you, it learns what does not, and stops wasting your time on it.

• Everything is local. Your memory, learned pathways, and failure history live on your Mac, not in a cloud profile.
• Master switches. Learned pathways and the anticipation suggestions each have a toggle you can turn off entirely.
• Reset clears it. An explicit reset wipes the learned memory; nothing else can.
• Memory never bypasses approval. Knowing your preferences makes the twin's drafts better; it does not let it send, buy, or post anything without your OK.

The trade is simple and yours to make: the more the twin learns, the more it acts like you, and you can dial that back or wipe it whenever you want. See Privacy & security for the full picture.