AI Agents With Memory You Can Edit

Why chat history is not memory

A long conversation with a model is a long conversation. What gets carried into the next reply is whatever fits in the window, which means the thing you said in week one competes with the thing you said this morning, and usually loses. You notice the failure a particular way: you tell it something important, and a month later it asks you the same question again.

The fix is not a bigger window. It is deciding what was worth keeping. After a conversation, an agent distils it into short standalone facts — the kind of sentence you would write on an index card — and those facts, not the transcript, are what get put in front of it next time.

One memory, many views

The important design choice is where the facts live. They belong to your account, not to the agent that happened to hear them. Ask your household agent what is on Thursday and ask your work agent the same question, and you get the same answer, because they are reading the same store.

What differs is the slice. Each agent is handed the facts that fit what it is for — a cooking agent does not need your meeting notes — and when a fact came from another agent it arrives attributed, so “how do you know that?” has a real answer instead of an uncanny one. Facts also carry their date, and a fact that supersedes an older one replaces it rather than piling on top, so nobody is still working from a plan you changed last week.

Except the things that stay with one agent

A shared memory is only acceptable if you can carve pieces out of it. So some facts are private to the agent you told, and privacy is enforced in the code at every exit — not requested in a prompt and hoped for. A private fact never crosses to another agent, never appears at a round table, and is never written into a summary, a hand-off, a briefing or an export.

Two things decide that a fact is private: you saying so, and a classifier that runs on every write regardless. The two signals are combined rather than trusted separately, because a model that forgets once has already broken the promise. And when you ask for something that would include a private fact — “summarise everything you know about me” — the agent asks first, rather than silently including it or silently leaving it out.

A glass box, not a black box

You can open any agent and read every fact it holds. Add one it never worked out. Correct one it got slightly wrong. Delete one you would rather it did not keep. Export the lot. There is no hidden second profile behind the visible list, which is the only version of “memory” worth trusting with your actual life.

What it changes day to day

• The agent that texts you first writes a message about your actual week, not a generic greeting.
• A daily check-in can ask about the specific task you said you would do — because it is on the record, with a date.
• A tutor knows which words you already have and recycles them until they stick.
• A new agent you hire on Friday already knows your name, your timezone and your habits, without an onboarding interview.

Which plans have it

All of them, including Free. Memory is not the upsell — the number of agents, the reply ceiling and the voice minutes are, and those are laid out honestly on pricing. Nothing is deleted when a plan ends either: your agents keep every memory whether you pay or not.