Building a useful agent · Memory
Memory and dreams: creating the next generation of AI assistants
July 9, 2026 · 6 min read
Getting access to ChatGPT felt like having an expert you could ask anything. But it was a new expert every time. You always started from the beginning. A good colleague or friend is rarely the one who knows the most; they are the one who knows you, your work, your situation. What they have that the expert doesn't is a shared history. Memories.
Most of us have too much going on. Work threads, the school calendar, the renovation, the insurance renewal: every domain wants a piece of your attention, and the pile follows you around.
It would be nice to let someone else handle some of it. AI assistants seem like a natural fit: excellent at answering questions, great at research. But you are still the one taking the task forward. However good the answers get, it is still a tool.
You don't hand real work to a tool. You hand it to someone you trust.
To really let go, you need trust: trust that it will actually be handled. Between people, that trust grows in relationships, and relationships are made of a shared history. Memories.
That is why memory has become one of the hottest problems in AI. OpenAI is already on the third version of ChatGPT's memory system, Dreaming, and every serious lab is working on the same thing.
From chat windows to dreams
The window
The first ChatGPT, and every assistant like it, remembered exactly one thing: a window of the current chat. The window kept growing with every new model, but the rule never changed: the assistant could only remember what was inside it. Everything else was gone.
You've felt this: the long chat that slowly loses the plot, the assistant that contradicts what you agreed on an hour ago. It wasn't being careless. The beginning of the conversation had fallen out of its world.
Keeping more of the conversation
By creating a summary of the oldest messages, more information fits in the window. Most providers use a version of this today.
It helps, but only for a while. As the conversation grows, summaries get summarized again, details quietly disappear, and the memory turns foggy. And the moment you notice the fog, trust starts to sink.
Memories and dreaming
So the modern systems added real memory: notes that persist across conversations. And the newest generation goes further, working on memory while you're away: a background process reads recent conversations, updates what it knows, and retires what's outdated.
In June, OpenAI shipped the third version of Dreaming, ChatGPT's memory system. The name is no coincidence, and that is the interesting part: as with so much in AI, the best model we have is the human one. Ours is modeled on it deliberately. Here is how.
Simulating human memory
Human memory is one of the best-mapped areas of psychology. The memories you can consciously call up come in two kinds: the things that happened to you, and the things you simply know.
Memories are created from experiences
Things you experienced, like the trip to Pisa, are captured by a part of your brain called the hippocampus and become episodic memories. They usually have a time, a place, and the people who were there.
It really does lean!Try recalling yesterday's lunch conversation word for word and you'll see why: your brain drops transcripts too. It keeps moments.
Memories carry different weight
Not every moment weighs the same, and the weighing happens immediately, not later. In your brain the amygdala tags experiences with emotional significance as they happen: it's why you'll remember the view from the top forever and the Baptistery line not at all.
Sorting the day into knowledge
While you sleep, your brain replays recent experience, groups what belongs together, and stores the repeated patterns in your neocortex as things you simply know, no longer tied to any single day. Our agents run this as a nightly pass: related episodes find each other and are distilled into semantic memories, durable knowledge with links back to every moment that supports it.
Notice what did not happen: the moment at the top didn't become a "fact". It doesn't need to. Some memories matter precisely as moments, and a system that understands that keeps them vivid instead of averaging them into a summary. Importance alone doesn't make knowledge; repetition does.
The right memory at the right moment
Every new message brings the right memories back. That is what makes an assistant feel like someone who knows you, not a stranger you brief from scratch each time. And it is what lets it give advice that fits you: the right suggestion for the right person, because it remembers who you are.
From a tool to a colleague
Remember where we started. A tool answers your questions and hands the task back to you. A colleague you trust takes the task and runs with it, because they know you, your work, and your situation. What separates the two is memory: the shared history a relationship is built on.
That is what everything above is for. Capturing experience, weighing it, sorting it into knowledge overnight, and pulling back the right piece at the right moment, so your assistant can build that history with you. Each part gets its own post, down to how we test it, including the time we caught our own test cheating. Memory is not a feature of the next assistant. It is what turns a tool into someone you can hand the job to.