Prompt InsightsSubscribe free
All terms

GlossaryUpdated

Agent memory

Agent memory is what an agent (AI that completes a task with several steps for you, not just one answer) saves outside the chat so it can remember things next time.


Agent memory is what an AI agent writes down outside the current chat, and reads back later, so it can remember you and your work next time. An agent is AI that completes a task with several steps for you, not just one answer. Memory is what stops it from starting from zero every time.

Here is the surprising part. The AI model itself remembers nothing between chats. Every new request starts blank. When an assistant "remembers" that you work in sales, or that your report goes to your manager every Friday, the app around the model saved that fact and quietly added it to your request.

What it means for you at work

Memory saves you from repeating yourself. You tell an assistant once that you write for customers in Germany, or that your team says "client" and never "customer". Next week, it still applies that.

Memory can also go wrong in quiet ways. If it saved something that is no longer true, it keeps using it. So treat memory like a shared notebook: useful, but worth checking.

Where you already see it

  • Chat apps. ChatGPT, Claude and Gemini can remember facts and preferences between chats when memory is turned on. You can usually see, edit and delete what was saved in the app's settings.
  • Projects and custom assistants. When you add files and instructions to a project, that is a kind of memory too. The app loads them every time you start a chat in that project.
  • Agents at work. An agent that sorts support tickets might keep a note of which customer is on which plan, so it does not have to look it up each time.

The four kinds of memory

Researchers often split AI memory into four kinds. The names come from the study of human memory:

  1. Working memory. What the AI is looking at right now: your request, the recent messages, the file you just shared. It disappears when the chat ends.
  2. Memory of events. What happened before: past chats, the steps of yesterday's task. Useful for "what did we try last time?"
  3. Facts. What is true about you and your work: your role, your preferences, your company's terms.
  4. How-to memory. Standing instructions on how to do a job, such as "always end a client email with clear next steps".

Many memory problems come from mixing these up. A one-off choice ("use blue for this slide") gets saved as a permanent preference ("always use blue").

How memory goes wrong

  • It gets old. A preference from March gets used in October, after things changed.
  • It contradicts itself. Two saved notes disagree, and the AI picks one without telling you.
  • It misses. The right note exists, but the AI does not find it. To you, it looks like it forgot.
  • It gets planted. If an agent saves what it reads, a hidden instruction in an email or web page can end up in its memory. That is a form of prompt injection, and it can come back in every later session.
  • It holds private data. Memory can keep more than you meant to share. Check what is saved before you use AI for sensitive work.

Memory and long chats are different things

In one long chat, an AI may shorten older messages to make room. This is called context compaction. It only happens inside one chat. Memory is what carries over to the next one.

The two meet in one place. If a summary of a long chat gets saved as memory, anything the summary got wrong gets saved too.

How to keep memory useful

  1. Look at what is saved. Open the memory settings once a month. Delete what is wrong or out of date.
  2. Say it plainly when you want something remembered. "Remember that our financial year starts in April" is clearer than hoping the AI picks it up.
  3. Keep rules in instructions, not in chat. Standing rules belong in custom instructions or project instructions, where they load every time.
  4. Use a temporary chat for one-off sensitive tasks. Many apps let you start a chat that does not use or update memory.

A prompt to review what AI remembers

Paste this at the start of a chat, in an app with memory turned on:

List everything you currently remember about me and my work. For each item, tell me:
1. what you remember, in one sentence
2. whether I told you directly or you worked it out
3. whether it might be out of date
Then suggest which items I should correct or delete.

The answer may not include everything, so also check the memory page in the app's settings. Then correct or delete what is wrong. Ten minutes a month keeps your assistant working from today's facts, not last year's.

Agent memory in the news

Questions

What is memory in an AI agent?

It is information the AI saves outside the current chat and reads back later: notes in a file, facts about you and your work, or a summary of past chats. The AI model itself remembers nothing between chats. The app around it stores the notes and adds them to your next request.

Does ChatGPT remember my previous chats?

It can, when memory is turned on. ChatGPT can save facts you share and use them in later chats, and you can see, edit or delete what it saved in its settings. Claude and Gemini offer similar features. Check what is saved now and then, because old facts can lead to wrong answers.

What are the types of agent memory?

A common way to split it has four kinds: working memory (what the AI is looking at right now), memory of events (what happened in earlier sessions), facts (your preferences, your company's terms) and how-to memory (standing instructions on how to do a job). The split comes from a 2023 research paper called CoALA, which borrowed it from the study of human memory.

Why does AI forget what I told it last week?

Because each new chat starts blank unless something was saved. If memory is off, or the fact was never saved, the AI has no way to know it. Put things you always want it to know in custom instructions or project instructions, which load every time.

What is the biggest risk with agent memory?

Wrong memories that stay. An old preference, a wrong fact, or a hidden instruction the AI picked up from an email or web page can be read back as true in every later session. Look at what is saved from time to time and delete what is wrong.

Related terms

Prompt packs for this

27,285 readers and counting

One new term, explained, in every issue