the infinite window

you open a new chat and explain yourself again. the plan, the thing you corrected yesterday, who "her" is. a long session does the opposite: it fills up, the oldest turns are summarised away, and something you said an hour ago is gone. both come from the same habit: treating the window an assistant can see as the place your memory lives.

the window is a cache

innernet treats it the other way round. your memory is the disk: large, kept, yours. the window is a cache: small, fast, chosen again on every turn. innernet is the pager in between, and a fast decider, a system one, chooses what moves into the window for this ask and what stays out. that decider, jev 1.14, answered at a p50 of 0.49 s when timed on its own.

scroll through one session to see what that means.

by design, a section the window already holds is never sent again: it is named by a handle. a section that changed arrives as a diff, the line that moved and not the page around it. when the window gets full it compacts, and anything compacted away comes back by its handle on the turn that needs it.

the session above is drawn: its asks and what arrives on each turn are illustrative. the numbers beside it are measured, on 90 asks from LoCoMo replayed as real chats: the window came to 7,437 tokens a turn, 80% of them read from the provider's cache, it compacted 9 times, and it re-sent 393 compacted memories in full when retrieval brought them back.

what it saves

the first test is LongMemEval-S: 100 questions over chat histories of 172,871 tokens each on average, read by Claude Haiku 5.5. sent the whole history, it got 90.0% right. through the window, bringing in 40 memories for each ask (k=40, fixed before the run), it got the same 90.0% right, on 25.9 times fewer tokens.

that is the part you would feel. an ask cost $0.00083 instead of $0.0870, about a hundred times less. give it fewer memories and it starts to miss: at 10 it fell to 81.0%.

two catches. the hundred times holds when the whole history is sent fresh each time; if a provider keeps it in its cache, the gap is an estimated 11 times. both are before ingest, cutting a history into memories, which costs $0.0287 a question and is paid once per history. and the window is not much quicker end to end, about 11%, because the expansion and the search before the answer give back most of the time the shorter answer saves.

where it lost

the second test is LoCoMo: three long conversations of about 20 to 40 thousand tokens each, replayed as real chats, 30 asks in a row. here the window lost. it answered 73.3% right, and the whole conversation, kept in cache, answered 87.8%. that is not noise (p 0.004). a conversation that size is cheap to send whole, and sending it whole is better.

so we put a switch between the two sizes we measured: while memory is small, send it whole and keep it cached. past about 100k tokens, the window. neither run locates 100k. it is where we put the switch, not a threshold the runs found, and the two runs differ in more than size: the benchmark, one ask against 30 in one chat, a history sent fresh against one kept in cache, and the judge's prompt. innernet's door is built to make that switch on its own, on a branch, and it is not yet live.

what it fixed

measuring the window meant measuring innernet's own door, and the door had two faults. the step that chose which sections answer an ask could only see the first 140 of a memory's sections. on innernet's own map of 1,551 sections, those held the right page for 2 of 22 test asks; the new shortlist holds it for all 22. and the step that picks the final sections had kept none of the ones it named on 149 calls in 3 days, because it copied whole lines where it was asked for names. after both fixes, an ask that found nothing in the morning found the right section later the same day.

what infinite means

bounded window, unbounded memory, flat cost. the test of that is whether success, the tokens each turn and the wait for the first word all stay flat as a memory grows from 10k to 10M tokens. that is the next measurement, on BEAM, and it is not in yet. the whole method, every table and what none of it shows are in white paper 02.

in a different setting, a small packet beat the whole history: in white paper 01, on 40 synthetic people, a person model's packet did 0.140 nats better than the whole history in context at 84.6 times fewer served tokens on the questions scored. two of its guardrails, G3 and G4, failed, and the people are programs. for that result, see what a person model knows.