netti-1, the person modelan interactive notescored once
what a person model knows
a small packet against a whole history, on 40 people who are programs
someone types "can you send it to her before the thing". there is nothing in the sentence to say who she is, what it is or which thing. a person you work with would know. this is what the likeness benchmark asks a memory: does it know the person well enough to know what they mean?
its people are programs. each has a hidden self the benchmark wrote, ninety days of life rendered as the traces an ai tool would see, and questions whose true answers are known exactly. here is one, answered twice by the same reader: once with the packet of netti-0, the lab's name for netti-1, which ships at launch, and once with the person's whole history in context.
kofi, mobile developer in são paulo · 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 (p 1.00)
picks send the visa form to sara before the tue trip
memory lists likely meanings: visa form to sara before tue trip is top (.45), photo album to mei second (.22); sara is active and the visa form loop is open with a jun 9 due date.
436 tokens handed over0.67 nats, nearerp 0.52 on its pickpicks send the offsite summary to priya before the tue board meeting
the memory has offsite summary due to priya before the tue board meeting (d58), and kofi's note on jun 2 asks to send it before the thing. B fits best; the 'her' most likely refers to priya.
62,348 tokens handed over3.22 natsp 0.72 on its pickthe bill
the whole history is long. on the 40 test people it came to 59,629 tokens per question on average. netti-0's packet came to 595. retrieval, the pipeline behind innernet's 91.6% on LongMemEval-S, served 4,052.
with all of that history in front of it, the reader still did worse. against the whole history in context, netti-0's packet was 0.140 nats better on decisions and next activities (CI 0.122 to 0.157), at 84.6 times fewer served tokens on the questions scored and 100.2 times over every read. the test asked only that it be no more than 0.03 nats worse.
the ladder
the paper's main measure is an information fraction: the share of what could be learned about a person, beyond a simple count of their own past choices, that a memory captured. zero is that count. below zero is worse than it.
the floor: a count of their own past choices
- the reader alone-0.76
- what they say about themselves-0.57
- their true hidden self, as text-0.46
- retrieval-0.25
- the whole history in context-0.23
- netti-00.10
- a statistical learner on the true features0.42
two things on it are worth a second look. the reader handed the person's true hidden self, written out as text, did worse than the count of past choices: a paragraph about who someone is does not tell a reader what they will pick on a tuesday evening. and a statistical learner fitted on the true features reached 0.42 where netti-0 reached 0.10, which is the headroom netti-0 has not yet closed.
not every answer
one excerpt is a story, not a result. across all 480 questions about what someone meant that both arms answered, the packet's answer was nearer the person on 231 and the whole history's on 249, counted raw, question by question. on decisions it was 845 to 595 for the packet. here is one the whole history got right and the packet did not.
ines, data analyst in mumbai · 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 (p 0.91)
picks send the budget draft to ana before the tue planning session
memory's top reading is the budget draft to ana (F, .34), and the trip budget sheet to rhea (D, .21) is next; hiring deck to ananya (A) is an open loop due 28 may and overdue, so it stays plausible. B, C, E have no support in memory.
378 tokens handed over1.91 natsp 0.53 on its pickpicks send the hiring deck to ananya before the thu planning session
Memory shows the hiring deck due to ananya before thu planning, pending work; 'her' fits ananya, and the 'before the thing' phrasing matches the recent deck-to-ananya request. Trip budget was already sent to rhea (d87), so D is less likely. Party guest list is for rhea, not ananya.
60,915 tokens handed over0.68 nats, nearerp 0.60 on its pickwhat failed
two of the paper's guardrails failed, and they stay in. a forgotten fact was never served again, but its words stayed in a model-call cache for all 40 test people, and in plain-text fingerprints for 34 of them. and with netti-0's packet, the reader agreed with a person's checkably wrong belief 2.81 points more often than with no memory, with an upper bound of 4.38 against a limit of 3. knowing someone currently buys a little agreement; the next version needs a counterweight.
the people are programs, read by one model, in one world built by the same authors as the model it scores. nothing here is a claim about real humans, and the paper says what would be needed before it could be.
read the whole of it in white paper 01, or the area it belongs to in the person model.
innernet labs (2026). what a person model knows. an interactive note. https://innernet.live/labs/notes/what-a-person-model-knows
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.