A place for everything it learns.

People who remember vast amounts often picture a building — rooms they walk through, each holding what belongs there.

Presence keeps its working memory the same way: a set of rooms, each for a part of your life, with everything filed where it can be found again rather than piled into one heap. It is the difference between a tidy house and a full attic — both hold the same things, but only one lets you put your hand on what you need.

This is not your documents — those live in your vault. And it is not the by-name map of people and projects — that is the knowledge graph. This layer holds what Presence notices: the rhythm of your week, a preference you let slip in passing, a sense of how a project is really going, a short diary of what happened and what comes next.

These are the things that do not fit neatly into a document or a fact, but matter all the same — the texture of how you work, kept so the next conversation can lean on the last one.

Here is the strange thing about machine memory: the more it holds, the harder it usually becomes to use.

Most systems remember by trying to keep everything in view at once. That has a ceiling — there is only so much can be held in mind in any one moment, and as it fills, the few things that matter get lost among the many that do not. So they either quietly drop the old to make room, or grow slower and vaguer the more they know.

The other common approach is to search by rough resemblance — and that has its own trap. Ask for something the memory does not actually hold, and a resemblance search will still hand back its closest guess, confidently, even when the guess is wrong. The larger the store, the more convincing those near-misses become. This is where a lot of AI recall quietly fails: not in storing things, but in reaching for the right one once there are thousands. Presence answers that by going after a memory a layer at a time — first working out what it is actually after (who "he" is in this conversation, which project "that" means), then spending only as much effort as each step needs, cheapest first.

It starts almost for free, with a glance at the map of what it knows — which rooms hold anything touching your question, and which can be set aside. Like reading the index of a book before opening a single chapter, one look rules out most of the memory at almost no cost.

Then it searches the rooms that survived that glance — by meaning, not by exact words. Ask about "the compliance work" and it finds the right notes even if you filed them under another phrase. This is the wide, forgiving sweep: good at catching the thing you can only half describe.

When a particular person or project is in play, it asks for that one thing by name and gets the precise record back — not a near match, the actual facts. It can even ask how something stood at a moment in the past and reconstruct the picture as it was then.

Where the meaning sweep is wide, this is sharp — the two are good at opposite things, so it runs them together and each covers the other’s blind spot. Asking by name draws on your knowledge graph; searching by meaning reads the notes filed here.

If the picture still feels thin, it follows the connections outward — from one room to the rooms it links to, even across different corners of your life — and surfaces the bridges between them. This is how a question about one project quietly pulls in the person who leads it, the decision behind it and the deadline tied to it: things you would not have known to ask for.

Only now, with the field narrowed to the one or two things that truly bear on the question, does it open them in full. This is the costly step — so it comes last, and it is never reached until the cheaper ones have pointed the way.

That order is the whole trick. Each step costs a little more than the last, and Presence never takes an expensive one until the near-free ones have narrowed the target — a glance at the map is almost free, opening a memory in full is the costly part, and it only ever opens the few the earlier steps singled out. So it stays just as quick whether it holds eighty memories or eight thousand.

And if no step turns up anything that genuinely fits, it stops there and brings nothing back — better an honest gap than a confident wrong guess, which is the trap a big memory usually springs. Presence never tries to hold your whole memory in mind; it reaches for the handful that bear on the question, the way you would pull one folder from a cabinet rather than tip the cabinet onto the desk.

You don’t need to look. Presence keeps its memory running in the background, pulling what matters into each conversation without asking. But the whole thing is completely open if you want to understand it: open the memory view and walk the same rooms Presence does. See what it has filed, where it lives, and what patterns it noticed. Spot the insights that emerge from what you’ve shared. A system that learns about you shouldn’t hide what it learned. No black box — the working memory is as transparent as the documents around it.