innernet labs
what we are finding out.
research on memory and the people it belongs to. six lines of it, each with the evidence it has so far, the limits written beside the numbers, and the failures left in.
- liveinnernet memoryone memory for every ai tool, kept current by netti and answered from it. live, and measured on LongMemEval-S.91.6%LongMemEval-S, all 500 questions
- tested insidenetti-1, the person modelwhat any ai can know of one person, and what it is worth to them, scored in nats per decision. tested inside; it ships at launch.84.6×fewer served tokens than the whole history, on the questions scored
- builtthe group modelhow a room thinks: why it exists, how it decides and what is still open, each item quoting the page it rests on.v0how we think, in every room
- a paperthe infinite windowa bounded context window filled from memory on every turn, measured against the whole history; its send-once ledger is not yet measured.25.9×fewer tokens than the whole history, same accuracy
- livethe resonance modela model of how a person thinks and works, read from how they answer, not from what they claim. you can take the quiz now.7 of 10checks passing on the last bench
- builtagent identityhow you want your agent to talk, answer and behave, and the habits it keeps for you, the same in every model.v0identities you can keep, one per agent or one for all
papers
written up in full, method and failures included.
notes
one result at a time, to read in a few minutes.
what it says
real answers, from people who are programs.
the same reader, the same question, answered once with the packet of netti-0, the lab's name for netti-1, and once with the whole history in context. picked to show both sides: on 480 questions about what someone meant, the packet was nearer on 231 and the history on 249.
kofi, mobile developer · a synthetic person
kofi (tue jun 2, 11:01) asks: "can you send it to her before the thing". which reading is right?
what the record says: send the visa form to sara before the tue trip
with netti-0’s packetpicks send the visa form to sara before the tue trip
436 tokens0.67 natswith the whole history in contextpicks send the offsite summary to priya before the tue board meeting
62,348 tokens3.22 natsines, data analyst · a synthetic person
ines (mon jun 1, 16:30) asks: "can you send it to her before the thing". which reading is right?
what the record says: send the hiring deck to ananya before the thu planning session
with netti-0’s packetpicks send the budget draft to ana before the tue planning session
378 tokens1.91 natswith the whole history in contextpicks send the hiring deck to ananya before the thu planning session
60,915 tokens0.68 natsjules, curriculum designer · a synthetic person
jules (tue jun 2, 16:27) asks: "follow up with him on that thing before the deadline". which reading is right?
what the record says: follow up with abel on the visa form before the fri dinner
with netti-0’s packetpicks follow up with abel on the visa form before the fri dinner
610 tokens0.42 natswith the whole history in contextpicks follow up with kwame on the quarterly deck before the sat board meeting
63,676 tokens2.30 nats
synthetic people only. nats are raw expected log loss against what an observer could know, computed per question for this page; the paper’s scored numbers are in section 6. more, typed out with their reasons, in what a person model knows.
how we work
said before, scored once, failures kept.
pre-registered
what counts as a result is written down and frozen before the test people are drawn.
scored once
the test split is scored a single time. anything run after that is labelled as such.
failures stay in
a guardrail that failed sits in the same table as the ones that held, with what comes next.
every number has a source
each one names the run or the file it came from. a number still being measured says measuring.
every benchmark we have run, and how each one was made: benchmarks. how our benchmark of person models works, and what it asks that recall does not: likeness.
a memory that knows where you are.
innernet keeps one memory for every ai tool you use, and netti keeps it current. this is the research behind it.