Protect · what the workspace knows
Two kinds of memory.
Both of them yours to erase.
LumenQube remembers in two separate ways, and they are not variations of one thing — different stores, different lifecycles, different rules. Settings → Memory & context shows both, shows the exact text they add to your next question, and erases either one. This page is what that screen is telling you.
What you told it, and what it worked out
The distinction matters because the two are stored differently, travel differently, and are forgotten differently. Software that calls both “memory” and shows you one switch is hiding a store from you.
Durable memory — what you told it
- Facts the assistant saved because you said them: preferences, projects, people, goals and everything else.
- Written only when something is durable. Credentials, keys, card numbers and recovery phrases are screened out before anything is saved or synced.
- It rides the cached part of the request, sent once rather than re‑sent with every question.
- Switch it off and the block is not built at all. There is no second copy elsewhere in the request — that was a real defect in the first draft of the context pack, and the review that caught it is why this page can say so plainly.
The context graph — what it connected
- Links between your own notes, tasks, meetings, events and people. Nothing is fetched to build it.
- It is drawn from what you already have, on this computer, in a local database.
- It travels as the context pack: titles and how far apart things are — never the contents. See the exact block on the homepage →
- Notes and tasks are held as sensitive and stay out of the plaintext index entirely, which is why the graph can reach an encrypted note that no search query could.
Where it is actually kept
Stated the way the code behaves rather than the way it would be tidier to describe.
The graph
This computer only. A local database beside your documents. It is not synced, not backed up to us, and not readable by anyone who has not unlocked this machine.
Durable memory
Two sets, and they do not reconcile. Up to 200 are kept here. The phone reads a separate cloud set of up to 64. A fact remembered on the desktop is not in the phone’s set, and one remembered there is not in these.
The screen says “this computer’s” for exactly that reason. Both caps are right for their own store; a surface that implied one memory across your devices would be the thing that is wrong.
Forget, and the five places it has to reach
Deleting a note is one row in one table. Forgetting it is five operations in four stores, and no cascade spans them — which is the whole reason this is a feature rather than a delete button.
-
The object
The record of the thing itself — the note, the task, the meeting.
-
Its links
Every edge to and from it. The database removes these on its own, which is why they are the easy half.
-
No cascade reaches here
The search index
Erased explicitly, because a cascade cannot reach a virtual table. Miss this step and the thing is gone while its title stays findable — the worst of both, and the kind of gap that only shows up when someone searches for what they just deleted.
-
A different process
The embeddings
Held in the app’s own working memory, in a different process from the database, so no amount of SQL reaches it. Before this shipped, forgetting a note left its text still scoring in semantic results until something else pushed it out — which on a quiet workspace is never.
-
Durable memories that name it
A separate store on a separate lifecycle, so it is asked separately rather than assumed.
Every store is attempted on its own, and you get a receipt for what was not reached. If the engine was switched off, or one store refused, the result says “Partly forgotten — some of it is still stored” and names the part. It does not show you a success. A Forget that quietly skipped a store is the exact failure the feature exists to prevent, so the partial state is the one that got its own review.
What this screen does not do
Listed because a privacy surface that only lists its features is not a privacy surface.
Not built
- No per‑space or per‑project AI‑access matrix. You cannot yet say “this project is never readable”. The control does not exist, so the words for it are not on the screen — a test checks they are absent, so the copy cannot arrive before the control.
- Memories carry no source links. You can read what was remembered, not the conversation it came from.
- The graph does not serve the search box. Quick search stays one round trip on purpose; the graph serves the assistant.
What it does instead
- Turn the engine on or off, with live counts by type and by where each thing came from.
- Show the exact text added to your next question — both blocks, verbatim, before you send anything.
- Forget what you are working on, across all five stores, with the receipt above.
- Say plainly where each store lives, including the two memory sets that do not agree.
Questions
Does LumenQube train on my documents?
No. Nothing you write is used to train a model, ours or anyone else’s. The graph is built on your own machine from files you already have, and the only thing that leaves is the request you asked for — which you can read in full before it goes.
What exactly gets sent when I ask a question?
Two blocks, and the screen shows both verbatim. The context pack carries titles and how far apart things are — never the contents of a file. The memory block carries the durable facts you asked it to keep, and is not built at all when memory is switched off. The homepage prints both, exactly as the app assembles them.
If I forget something, is it really gone?
From LumenQube, yes — across the object, its links, the search index, the embeddings, and any memory naming it. If a store cannot be reached, you are told which one rather than shown a success. What Forget does not do is reach a copy you made somewhere else, or a message you already sent to a model in a previous conversation.
Why is memory on my phone different?
Because they are two stores, not one view of a store. The desktop keeps up to 200 locally; the phone reads a cloud set of up to 64. They never reconcile, so the two can genuinely disagree. That is stated on the screen rather than smoothed over, because a memory you cannot find is worse than a memory you were told about.
Can I turn all of this off?
Yes, and separately. Memory off means the memory block is never built. The engine off means no context pack is assembled. Neither switch degrades the editors — ordinary work does not consult either, and it keeps working with the network off entirely.
The rest of the boundary
The exact block, printedHow everything is securedKnow — find your own pastCalendar — excluded means never asked