# a memory that knows the person

> netti-1, the memory our lab measured, is now what every innernet runs on: what your ais are handed, what your agents recall, and a reading of you that netti folds in each night.

by jonna · news · 2026-11-18

https://innernet.live/company/a-memory-that-knows-the-person

---

most memories for ai keep sentences. you said you prefer short answers, it is written down, and the next time a question shares a word with it, the sentence comes back. that is useful, and it is where most of the field has settled: on the usual recall test, the published scores now sit within a few points of each other.

what a store of sentences cannot tell an ai is who it is talking to. what you mean by what you ask. what you would choose. what changed about you last month, and what was a one-off. netti-1 is our answer to that. it came out of our lab, and it is now what every innernet runs on, on every plan.

## four parts, one memory

netti-1 is not a new model you talk to. it is the way your memory is read, and it has four parts.

**what your ais are handed.** when claude, chatgpt or cursor asks innernet about you, it used to get the few facts that shared a word with the question. now every fact that app is allowed to read is weighed against the ask, and the ones that bear on it come back best first, inside a small budget. a fact that rests on something you said carries your own words beside it.

**what a project hands over.** a memory that grows cannot be sent whole on every question. so on each ask, innernet chooses the sections of your own project that answer it and sends only those, with the rest named so the ai can go deeper if the conversation turns. your context stays small however large the memory gets.

**what your agents recall.** an agent you build keeps what was said beside what it learned from it, and a recall reads both. this is the part an agent's answers lean on, and the part we measured hardest.

**a reading of you.** each night, after your own midnight, netti folds the day you had into a reading of how you work: your own words, what your ais noted, what you did. one day at a time, in order. it holds tendencies, not rules. when an ai asks about you the next day, the part of that reading that bears on the question rides along, with no extra model call.

## what stays yours

the reading is built from everything, including what you keep private, so it follows a stricter rule than a single fact does. it is kept sealed under your own key. it rides only for an app you have cleared for every class of memory it was read from, and when anything is in doubt it does not ride at all. when you forget something, the forget is carried into the reading before anything else is read, and no reading is served while one is waiting.

choosing which sections answer a question is done by a small decision model, called with zero data retention asked on every request. our [privacy page](https://innernet.live/privacy) names who runs it and what it is sent. if that call cannot be made the way we ask, it is not sent anywhere else, and innernet carries on without it.

and there is one switch above all of it. any part can be closed on its own, and when a part is closed your memory answers exactly as it did before.

## what we measured

we would rather show the numbers than describe them.

- **89.4% on LongMemEval-S.** this is innernet's agent memory on the same code your agents call: 447 of all 500 questions, and 313 of the 350 our tuning never saw. Claude Haiku 5.5 read the answers and Claude Sonnet 5.5 graded them with the test's own judge prompts, not gpt-4o, so it does not line up exactly with scores graded another way. its weakest kind of question is a single-session preference, at 76.7%.
- **92.0% is the research build, not the product.** it is the memory netti-1 grew from. it reads about two and a half times the tokens, and it is never a memory anyone connects to. we say so wherever the number appears.
- **what your ais read about you is a different door, and it scores lower on purpose.** on a year of chats from MemoryAgentBench, it answered 17.0% before netti-1. with netti-1 it answers 31.7% for a standard app, and 60.3% for an app you have cleared for what is personal. the distance between those two numbers is your privacy settings doing their work: a recall test cannot reward a memory for what it rightly holds back.
- **the person model itself** was scored once, after a freeze, on 40 synthetic people. its reading of a person ran to about 600 tokens, against about 60,000 for their whole history, and it predicted what they would choose better than the whole history did. the people are programs, so this is not a claim about real humans, and two of its guardrails failed. the paper keeps both failures in.

every one of these, with how it was run and where it lost, is in [innernet labs](https://innernet.live/labs): [the person model](https://innernet.live/labs/person-model), its [white paper](https://innernet.live/labs/papers/person-models) and a shorter [note on what a person model knows](https://innernet.live/labs/notes/what-a-person-model-knows), [likeness](https://innernet.live/labs/likeness), the benchmark we built because recall tests stopped telling memories apart, [the benchmarks board](https://innernet.live/labs/benchmarks), and [the infinite window](https://innernet.live/labs/infinite-window), the research behind keeping your context small.

## if you build

netti-1 is what an agent's memory runs on too. two calls a turn give any agent a memory of its own, and [an agent's memory belongs to someone](https://innernet.live/company/an-agents-memory-belongs-to-someone) says how ours is put together. the [developers page](https://innernet.live/developers) has the block to paste, the [agents guide](https://innernet.live/docs/agents) and the [api reference](https://innernet.live/docs/agents-api) have every call, and [pricing](https://innernet.live/pricing) is one page: reading is free on every plan.

## if you just use it

there is nothing to turn on. the ais you have connected are already reading your memory this way. if you would like to see what netti holds about you, and what each app may read, it is all on [your page](https://innernet.live/dashboard/you). and if you have not started yet, [you are welcome to](https://innernet.live/signup).
