What is AI memory? Explained with an assistant's notebook
AI memory is information an AI saves outside the conversation and reads back next time, because the model starts each chat knowing nothing. Explained with an assistant who keeps a notebook between meetings.
- Author
Umair Mateen- Published
- Updated
- Reading time
- 5 min
- Category
- AI
On 6 November, Anthropic published research on how AI agents remember things between sessions, so here is what "AI memory" actually means.
In one sentence: AI memory is information an AI assistant saves outside the conversation, such as notes about your preferences or lessons from earlier work, and reads back in a later conversation, because the model itself starts every new conversation knowing nothing about you.
The everyday picture
Imagine a very capable office assistant with one odd habit. Every morning, they arrive with no memory of yesterday. Whatever was said in Monday's meeting is gone by Tuesday.
So the assistant keeps a notebook. During each meeting, they jot down what matters: "The client prefers short reports." "The invoice template moved to the shared drive." Before the next meeting, they open the notebook on their desk and read the useful pages.
The assistant is the AI model. The desk is the context window, the limited space the AI can look at during one conversation. The notebook is the memory.
The assistant forgets each meeting once it ends. What it wrote in its notebook is put back on the desk next time, so the next answer can use it.
How it works, step by step
- A conversation happens (Monday's meeting). You tell the AI something worth keeping, or it learns something by doing a task, such as which approach failed.
- The AI writes a note (a line in the notebook). The note is saved as plain information in storage outside the model. Anthropic describes agent memory as knowledge the AI keeps between sessions, "saved as context it can read later". The model's own training does not change.
- The notes are kept in layers (the notebook has sections). Anthropic's research describes full records of past sessions that can be searched, short notes the AI writes, tidier summaries such as a wiki, and a small top layer that is loaded into every new session.
- A new conversation starts (Tuesday's meeting). The small top layer is copied onto the desk straight away. Deeper notes are only looked up when the task needs them, which keeps the desk clear.
- The AI answers with the notes in view (the assistant who checked the notebook). It can skip questions you already answered and avoid mistakes it made before.
Some systems also tidy the notebook between meetings. Anthropic says this idle-time review, often called "dreaming", can fix mistakes and write better notes, but it costs extra computing.
Try it · Put it in order
How does an AI remember between conversations?
Tap the steps in the order they happen.
Where the picture stops
- A person remembers some things without trying. An AI keeps only what is written down and read back. If nothing is saved, nothing carries over.
- A notebook is private to one assistant. AI memory is stored data in a product or in a company's own system, so who can see it, and how long it is kept, depends on how that system is built.
- A person may doubt an old note. An AI tends to trust what it reads. Anthropic warns that a note that is wrong, or goes out of date, can be read and trusted in a later session.
Why it matters for your business
- Less repetition: staff stop re-explaining the same background, formats and preferences in every chat.
- Better work over time: in Anthropic's own test, where an agent answered database questions across fresh sessions, every Claude model it tried scored higher with memory than without.
- New risks to manage: a wrong note can be carried into every future session, and notes loaded each time take up space on the desk, which adds cost.
- Data control: memory is stored information. With Anthropic's memory tool for developers, the files live in your own application, so your team decides where they are kept and when they are deleted.
An example
Imagine a 20-person accounting firm using an AI assistant to draft monthly client summaries. In January, a partner tells it that one client wants figures in euros and no more than one page. The assistant saves a note. In February, a different staff member asks for that client's summary, and the assistant follows both rules without being told. In March, the client asks for two pages. Someone has to update the note, or the assistant will keep following the old rule.
Frequently asked questions
Is AI memory the same as the context window?
No. The context window is what the AI can see during one conversation. Memory is stored outside it and copied back in when needed, so it can survive after a conversation ends.
Does the AI learn permanently from my conversations?
Not through memory. Memory notes are saved data that the model reads. The model itself is not retrained, and a deleted note can no longer be read back.
Can AI memory be wrong?
Yes. A note can be mistaken from the start or become outdated as things change. That is why it helps when people can view, correct and delete what has been saved.
Is memory the same as summarising a long chat?
No. Summarising, sometimes called compaction, shortens one conversation so it fits on the desk. Anthropic treats memory as notes kept outside the context window and pulled back in at later times.
Related explainers
- What is a context window? Explained with a clerk's desk
- What is an AI agent? Explained with an assistant you send on errands
- What are tokens in AI? Explained with word-game tiles
- What are AI hallucinations? Explained with a confident quiz guess
The DevDrift builds AI assistants with memory your team can review and control: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.