the infinite window

a bounded context window filled from memory on every turn, measured against the whole history; its send-once ledger is not yet measured.

what it is

a context window is a cache. memory is the disk, innernet is the pager, and a fast system one decides what moves between them on every turn.

by design, the packet is bounded by the ask, a thing already held is referred to by a handle and not sent again, a change arrives as a diff, the prefix stays the same so the provider’s cache keeps it, and compaction is only the backstop. what is built is less: handles are labels, a compacted memory is sent again only when retrieval brings it back, and diffs and a reader that asks for a handle are designed, not built.

white paper 02 measures it. on LongMemEval-S the window kept the whole history’s accuracy at k=40 on 25.9 times fewer task tokens. on LoCoMo, where a conversation is a few tens of thousands of tokens, it lost to the whole conversation kept in cache. between those two sizes we put a switch: while memory is small, send it whole and keep it cached; past about 100k tokens, the bounded window. neither run locates 100k: it is where the switch was put between the two sizes measured, not a threshold either run found, and the two runs differ in more than size.

the evidence

LongMemEval-S, the window at k=40 against the whole history25.9× fewer task tokens, 90.0% each. 6,675 task tokens a question against 173,198, paired +2 / -2, exact McNemar p 1.000, on 100 held-out questions read by Claude Haiku 5.5packages/bench/results-infinite-window-lmes-haiku-5-5.json (PR #275), 2026-10-16
what an ask costs at k=40$0.00083 against $0.0870. the whole history sent fresh on every ask, about a hundred times; kept in a provider cache it is an estimated $0.00946 an ask after a $0.1089 first write, so the gap narrows to about 11 times. both are before ingest, $0.0287 a question, paid once per historypackages/bench/results-infinite-window-report.md (PR #275), 2026-10-16
LongMemEval-S, the window at k=1081.0%, 72.6× fewer task tokens. too few memories loses answers: +3 / -12 against the whole history, exact McNemar p 0.035packages/bench/results-infinite-window-lmes-haiku-5-5.json (PR #275), 2026-10-16
LoCoMo, 90 asks replayed as true multi-turn chats73.3% against 87.8%, the window lost. a whole conversation kept in cache beat the window (+3 / -16, p 0.004) at 33,819 prompt tokens a turn against 7,437, on conversations of about 20,079 to 40,139 tokenspackages/bench/results-infinite-window-locomo-session.json (PR #275), 2026-10-16
the door's shortlist, 22 labelled asks on a map of 1,551 sectionsa gold page for 22 of 22 asks, from 2 of 22. the BM25 shortlist's top 60 over every section, against the first 140 sections in load order that the read used to seecontextbench recall@N on innernet's own map (PR #272), 2026-10-16
the door's llm selectorkept 3 of 3 sections it named, from 0 of 2. it had kept no section it named on any of 149 production calls in 3 days; short ids fixed it, at 22 output tokens instead of 114live A/B on Claude Haiku 5.5 (PR #274), 2026-10-16
the door, live on innernet.livean ask that found nothing in the morning found the right section. 2 focus sections, about 363 tokens, chosen from the 1,591 sections the live read considered that day; a later ask ran the llm selector: 3 focus sections, about 1,566 focus tokens, the select step 2.1 sinnernet_context's own sent line, after PRs #272 and #274, 2026-10-16
the decider, jev 1.14 through OpenRouterp50 0.49 s. 100 of 100 calls at about 1,160 a minute with no throttling, $0.042 per million input tokens, zero retention per requesta load run against the decider alone, outside the door, 2026-10-15 to 2026-10-16
ask netti, the first wordp50 2.29 s, from 4.23 s. 51 cases on Claude Haiku 5.5, cold; 50 of 51 passing, from 48 of 51; with jev live the first word was 2.63 sthe ask netti bench (PR #276), 2026-10-16
the desk's hands, with jev beside Haiku0.64 steps a second against 0.62, 18 of 18 each. on a fake desk with modelled hand time, six click tasks three times; $0.0009 a run against $0.0014; ships darkthe desk click bench (PR #277), 2026-10-16

100 questions: every 95% interval is 12 to 15 points wide, so a small gap between arms is not a finding
one reader, Claude Haiku 5.5; the k values were fixed before the run, not tuned on it
the ask path is only about 11% faster end to end: the expansion and the search give back most of what the shorter answer call saves
a whole history kept in a provider cache narrows the dollar gap to an estimated 11 times, and both gaps are before ingest, $0.0287 a question paid once per history
on LoCoMo, three conversations of about 20 to 40 thousand tokens, the window lost to the whole conversation kept in cache; that is one run on one benchmark, not a rule for every memory that size
the desk numbers are on a fake desk with modelled hand time, and jev did not bring ask netti’s first word sooner: 2.63 s against 2.29 s

papers and notes