Your assistant forgets you every session. This one doesn't.

Human memory does not have one store. It has a fast one that records today and a slow one that keeps what matters, and the fast one hands over to the slow one during sleep — the complementary-systems account that McClelland, McNaughton and O’Reilly set out in 1995, building on Marr’s 1971 theory of the hippocampus. That is the shape of this engine: everything lands in the fast store immediately, and a nightly pass moves the durable parts across. The rest of the night’s work comes from the same field — replaying related episodes together, as Wilson and McNaughton observed happening during sleep; reopening an old memory when something new touches it, which is Nader and LeDoux’s reconsolidation; and idle-time reflection, the default mode Raichle described.

Memories are not a list, they are a graph, and a graph has neighbourhoods. Finding them is a studied problem, and the modern standard is the Leiden method of Traag, Waltman and van Eck (2019), scored with the CPM objective from Traag’s earlier 2011 work. That objective is the piece of published science we have in common with them, and the credit for it is theirs.

What clusters your memory every night is not that algorithm. It is ours, and it is a different machine — because our problem is one the standard was never asked to solve. Three differences carry the weight:

Communities keep their identity. Run ordinary clustering twice on a slightly changed graph and you get two unrelated sets of groups with fresh arbitrary labels. Ours tracks every community across every nightly pass and through every split and merge, so “the pricing cluster” is still the same thing next week and its history is something you can ask about. Standard implementations have no notion of this at all.

It settles its own resolution. How coarse or fine the grouping should be is normally a number a human picks, and re-picks whenever the data shifts. Ours decides that itself against several competing goals at once, every cycle — because nobody is going to hand-tune a parameter on a laptop at three in the morning.

It is built to run out of time. Clustering here has a hard wall-clock budget: it gets the idle window and not a second more, and it has to hand back a good grouping when that window closes rather than a perfect one eventually. That single constraint shapes the whole design, and offline graph tooling does not have it.

Same question as theirs, same scoring function, an entirely different answer to how you compute it under these conditions. We name them because that is how honest engineering reads — you say whose shoulders you stand on, then you say plainly what you built yourself.

Pentti Kanerva’s sparse distributed memory, and the wider vector-symbolic tradition that Tony Plate and Ross Gayler shaped, represent things as very wide patterns you can combine and compare whole. It is why a moment, a piece of knowledge and a habit can share one store without collapsing into the same mush, and why you can find a memory by its shape and not only by its words.

Not every gap in what it knows is worth asking about. Shannon gave us the measure of how much an answer would actually tell you, and the engine uses it to decide which few questions are worth keeping. Without that measure, a system that notices its own gaps would bury you in them.

An old and unfashionable field that got there first. From W. Ross Ashby comes the idea that a system stays alive by holding a few essential variables inside survivable bounds — which is exactly what the engine watches while it reorganises itself, rather than optimising one number into the ground. From Gordon Pask comes something we use almost literally: his teach-back, where understanding is demonstrated by explaining a thing back and being corrected where the explanation fails. After each consolidation the engine attempts precisely that on its own conclusions and flags what does not hold together. Wiener named the field in 1948; it deserves more than nostalgia.

This is where the name comes from, and it is a design decision, not a metaphor. Monotropism — the account of attention Dinah Murray, Mike Lesser and Wenn Lawson published in 2005 — describes deep narrow focus rather than shallow broad attention, and that is how recall here behaves. The exact words instead of the gist. Rare things kept rare instead of averaged away. Most memory systems are built to generalise; this one is built to be literal — which is precisely what you want from something that has to remember what you actually said.