Results · the reading chain

Agents in a chain: at the same size, a memory handoff keeps twice what a written note keeps

Every agent is Qwen3-4B, on one H100 GPU. Eight copies stand in a line: each reads one Wikipedia paragraph and hands on to the next; the last answers questions on all eight paragraphs. The agents in the middle never see the questions, so the handoff must carry everything. The handoff is a written note or the sender's working memory cut to the same size, at six sizes from 64 to 2,048.

The ladder: every size of handoff

At every size the cut memory keeps at least what the note keeps; from 512 it keeps twice as much, and at 2,048 everything

The handoff's size is the budget: tokens for the note, positions for the cut memory, the same number for both. The whole memory and one agent reading everything are the ceiling. The score is token F1, the overlap between the answer given and the right one, over every question of the eight documents; eight documents are about 1,300 tokens, so from 2,048 the budget holds everything.

Every rung at every model size. Two rules pick which positions of the memory stay; neither sees the questions; the second is the better one. Three seeds a cell. At 256 and under the first rule loses to the note, and a plain cut of each document's first lines, with no model at all, beats both rules there.

What each rung costs. The same scores against the tokens each chain ran. A note is written and re-read at every hop, so its cost grows with the budget; a cut memory is not. From 512 on the cut memory keeps more for less.

Three handoffs

What crosses: words, the model's memory, or that memory cut to the words' size

Written noteeach agent reads its document and writes a note of at most 256 tokens that must keep every fact so far; the next agent reads the note

Whole memorynothing is written: the next agent continues the sender's working memory as if it had read every document itself, and reads its own on top

Memory cut to sizethe same memory, cut to 256 positions, the room the note gets; a rule that never sees the questions picks which positions stay

	Written note	Whole memory	Memory cut to size
What crosses	the sender's description of what it read, in words	the sender's working memory as it stands, nothing written	the same memory, cut to the note's number of positions
Size at hop 8	tokens, about	positions,	positions,
Who decides what is kept	the writer, under its word budget	nobody: nothing is dropped	a rule that never sees the questions
What a hop costs	the sender writes up to 256 tokens one at a time: s a hop	a copy; the receiver only reads its own document: s a hop	a copy and the cut: s a hop

One chain, hop by hop, at the same size

At a hop: under a written note a document fades from the hop after it is read; under the cut memory it stays lit

Eight documents, one a hop; the handoff tokens or positions Hop 0of 8

A tile is one document; its brightness is the share of its questions still answered right after that hop. The note and the cut memory get the same room. The bar at the right is the score over every question so far; the dotted tick is one agent reading everything itself.

Written noteCut memory, first ruleCut memory, better ruleWhole memoryOne agent reads everythingClosed book: no document at all

Metrics, at the same size

The cut memory against the written note at a hop, after seven handoffs

Compute is every token the model ran a forward pass over. Bytes are what one hop moves: four a token for a note, KB a position for a memory at this model size. The cut memory wins on facts and time and pays in bytes.

The result in one picture

Score after each hop at : the note falls from the first handoff, the cut memory holds for the first hops and fades slowly

The score is over every question of every document read so far. The shaded band is what the note loses to the cut memory; the dotted floor is a model given no document at all.

Which facts, and when

Most of what a note loses dies at the first rewriting; the cut memory loses slowly and evenly

When. Of the facts still alive before a handoff, the share lost at it. The note loses most at the first rewriting; what survives it is what every later writer copies. The cut memory loses a few per cent a hop.

Which. Facts kept after seven handoffs by the kind of answer. A writer under a budget lists numbers and names and drops the phrases it would have to paraphrase; the cut memory keeps every kind alike.

The gap, with its uncertainty, at every model size

The cut memory never loses to the note, at any size of model or handoff; the gap grows with the room

Paired on the chain: the same eight documents in the same order under both handoffs, the difference averaged over every chain and seed, with its 95% interval. An interval wholly right of zero is a win. The better rule is level with the note at the two smallest sizes and ahead everywhere else; a bigger writer does not close the gap.