# an agent's memory belongs to someone

> memory for the agents you build: two calls a turn, a page you can read for every agent, and what it learns about you waits for your yes.

by jonna · news · 2026-11-18

https://innernet.live/company/an-agents-memory-belongs-to-someone

---

an agent that forgets is easy to build. it answers well, the chat ends, and the next person who writes in is a stranger again, even when it is the same person. so most of us reach for a memory, and most memories for agents are a place to put text and a way to search it. the agent gets a drawer. nobody can see into the drawer, the person it is about has no say in it, and when something in it is wrong it stays wrong.

we built agent memory in innernet from a different starting point: every agent belongs to someone, and so does everything it remembers.

## two calls a turn

there is no sdk to learn. before your agent answers, it asks innernet for context, with the person's id and their message, and puts what comes back in the prompt. after the reply has gone out, it hands innernet the exchange. that is the whole integration: plain https and json, from any language, in front of any model.

you do not even have to write those two calls. make an agent, paste one block into the ai you code with, and it reads [the guide written for it](https://innernet.live/skills/agent-memory), finds where your agent answers, keeps the key in an environment variable and adds both calls. or ask the ai you already talk to, and it does the whole thing from the chat. or use one line in a terminal.

## it decides what is worth keeping

this is not a vector database with a new name. your agent hands over the exchange, and innernet decides what in it is worth keeping. what it keeps is sorted by what kind of thing it is: how the agent works, what it knows, what people like, what happened, and notes. each memory says who it is about. it is versioned. and it is forgotten on a schedule you set: by default something that happened once fades after 90 days, while facts and how-tos stay until you let them go.

recall ranks by words, recency and importance, and can re-read the best matches by meaning. the sorting happens on our side, so your agent makes no extra model calls of its own.

## who a memory is about

every memory an agent keeps is about one of three: the agent itself, one of the people it serves, or you, the person the agent belongs to.

what it knows about a person it serves stays with the agent. every call carries that person's id, and one person's history never reaches another person's answer. when someone asks what your agent knows about them, you can show them. a dry run says exactly what would go, and then one call forgets that person entirely, every memory and every version.

what it learns about you is the part we care about most. it does not quietly become part of your record. by default it waits, and you keep it or let it go. if you keep it, it arrives in the agent's voice, marked as a note an agent made. it can never read as something you said, and it can never reword something you did say. something you let go of stays let go of, even when the agent hears it again.

## one level, and a page you can read

how far an agent may reach into your own memory is one choice. sealed reads nothing of you. asks first, the default, reads what any connected app may and holds what it learns for your yes. trusted reads your personal memory and files what it learns on its own. a key or a chat can narrow that by itself. widening it takes you, signed in, saying yes.

and you can see all of it. every agent gets a page of its own: what it keeps, on its shelves; what it may read and write of you; everything it did, by day; and how it remembers. you can change any of it without touching code. an agent that remembers is only safe to hand real work to if you can see and change what it remembers.

## contained

an agent's key reaches that agent and nothing else: not another agent, not your projects, not your own memory directly. keys and card numbers are taken out before anything is stored. every memory comes out as json whenever you ask. remove an agent and its memory, its keys and its record go with it, while what you chose to keep stays yours. what your agents keep never trains a third-party model, and we never train our own on identifiable memory.

## one memory, not another one

each agent is made with a project of its own, and what it knows about its job flows into it as pages you can read. so an agent's memory is part of the same memory you already carry between claude, chatgpt and cursor. the ai you write code with can find what your support agent knows about its job, as far as you allow.

## measured, on the code you call

on LongMemEval-S, a public test of memory over long conversations, this memory answered 89.4% of all 500 questions, running the same code your agents call. Claude Haiku 5.5 read the answers and Claude Sonnet 5.5 graded them with the test's own judge prompts rather than gpt-4o. a research build of ours scores 92.0% on more than twice the reading. that build is not what you connect to, so 89.4 is the number we give. how it works underneath is in [a memory that knows the person](https://innernet.live/company/a-memory-that-knows-the-person), and every run is in [innernet labs](https://innernet.live/labs/memory).

## what it costs, and where to start

reading never counts: context and recall are free on every plan. only a fold counts, one pass of netti filing what was kept, and an agent's keep is a fold like yours. 200 a month are free, and plus is $14 a month for as many as you need. one account holds up to 100 agents, each keeping 10,000 memories.

it is open today. [the developers page](https://innernet.live/developers) has the block to paste, [the agents guide](https://innernet.live/docs/agents) walks through the first one, and [the api reference](https://innernet.live/docs/agents-api) has every call. if you are weighing it against something else, [the comparisons](https://innernet.live/company/compare) say where each one fits. and if something is missing for what you are building, please tell us at connect@innernet.live.
