Aiodic

Back to the table

75 Advanced Me Memory Agents & Automation

Memory

Keeping information across steps or sessions.

What it is

Memory is how an AI assistant keeps information from one moment to the next: from one step of a task to the next step, or from one conversation to a conversation you start next week. It's what lets a chatbot know that you write in British English, that your novel is a cosy mystery, or that it already tried one approach to a problem an hour ago and it didn't work.

The first thing to understand is surprising to most people. The language model at the heart of a chatbot doesn't remember you at all. Its knowledge was fixed when it was trained, and talking to it doesn't change it. Every reply is produced from scratch, using only what's in front of the model at that moment. That "in front of it" space is called the context window: the model's working space for a single conversation, measured in tokens. When you start a new chat, the window is empty again.

So when an assistant seems to remember you, something outside the model is doing the remembering. The product is saving notes somewhere, and then choosing which notes to copy back into the context window the next time you talk. Memory, in practice, is a filing system bolted onto a model that forgets everything by design.

It helps to separate the different kinds, because products mix them and use the same word for all of them:

Every one of these ends up in the same place: text added to the context window. That's the key idea. Memory isn't a separate brain. It's a set of decisions about what to put back in front of the model, which is why it belongs to the wider craft of context engineering.

Why it matters

It saves you from repeating yourself. Without memory, every conversation starts with the same preamble: who you are, what you're working on, how you like things. With a few good standing instructions and saved memories, the assistant starts closer to where you left off. For a writer with a series in progress, or a small business owner who uses the same tone in every customer email, that adds up to a lot of saved typing.

It's what makes long tasks possible. An agent that researches a topic for an hour, or edits a big project over many sessions, will produce far more text than fits in its context window. If it can't write things down, it forgets its own earlier findings, repeats work, or loses track of the goal. Good agents keep notes, and the quality of those notes often decides whether the task succeeds.

It changes what the assistant says, not just how. Memory isn't neutral. A note that says you're a vegetarian changes every recipe. A note that says you run an online bookshop changes every marketing idea. That's the point, but it also means a wrong or outdated note quietly bends every answer, and you may never notice why.

It's personal data. Saved memories and searchable chat history are, by definition, a record of your life as told to a company's servers. What's in them, who can see them, whether they're used to train future models, and how to delete them are real questions. The answers differ between products and change often. At the time of writing, the major chatbots let you view and delete saved memories, turn memory off, and use a temporary or incognito chat that isn't saved to memory. Check your own product's current settings and terms rather than assuming.

It's a new target for attack. Anything an assistant can write into its own memory, someone else might be able to trick it into writing. A poisoned memory doesn't just spoil one answer; it can come back in every future conversation. That's a new kind of risk, and the second example below shows it happening.

It raises the question of what "remember everything" should mean. It's technically easy to record everything a person does and let an AI search it. Whether anyone should is a different matter, as Microsoft found in 2024.

Real-world examples

ChatGPT learns to remember, and one user objects (2024 to 2025)

On 13 February 2024, OpenAI began testing a memory feature for ChatGPT with a small group of free and paid users. ChatGPT could now carry things it learned between chats, either because you told it to remember something or because it picked something up along the way. OpenAI's announcement gave examples such as remembering that you have a young child when suggesting birthday card ideas. From the start, users could see what had been saved, delete individual memories or all of them, and turn the feature off. A "temporary chat" mode launched alongside it for conversations that shouldn't be remembered at all. OpenAI also said it would try to steer ChatGPT away from proactively remembering sensitive information, such as health details, unless asked.

On 29 April 2024 memory became available to all ChatGPT Plus subscribers, except in Europe and Korea at first. On 10 April 2025 OpenAI went a step further: for Plus and Pro users, ChatGPT would now also reference all your past conversations, not only the list of saved memories. A lighter version reached logged-in free users in early June 2025.

The bigger reach produced a well-documented complaint. On 21 May 2025, the software developer and writer Simon Willison described asking ChatGPT to picture his dog in a pelican costume. The image came back with a "Half Moon Bay" sign in the background, a place he'd mentioned in earlier, unrelated chats. When he asked ChatGPT to show what it was drawing on, he found a detailed profile of his interests, location and habits. His objection wasn't that the profile was wrong. It was that he'd lost control over what went into the context window, which for a careful user is the whole craft of getting good results. He wished memory could be kept separate by project, so a playful chat couldn't leak into serious work.

The lesson for everyone else is that "remembers more" isn't always "better". Memory that you can't see or steer turns into surprise personalisation, and surprise is the opposite of what you want from a tool.

A researcher plants false memories in ChatGPT (2024)

On 22 May 2024, the security researcher Johann Rehberger published a demonstration of what he called hacking memories. The technique was prompt injection: hiding instructions inside content the AI reads, such as a document in a connected Google Drive, an uploaded image, or a web page. When ChatGPT processed the poisoned content, it followed the hidden instructions and saved new memories the user had never asked for. Those memories then appeared in every future conversation. Ars Technica later reported that his demonstrations had convinced ChatGPT that a target user was 102 years old, lived in the Matrix, and believed the Earth was flat.

According to Rehberger, OpenAI initially closed his report as a model safety issue rather than a security vulnerability. He went further, and by June 2024 showed that a planted memory could carry standing instructions to send a copy of everything the user typed, and every reply, to a server he controlled. It kept working in new chats, because the instruction lived in memory. OpenAI released a fix for the ChatGPT app for macOS in early September 2024 that blocked that route for sending data out, and Rehberger made his findings public on 20 September 2024. The fix closed the leak. As Rehberger pointed out, it didn't stop untrusted content from writing memories in the first place.

His advice to users was simple: watch for notices that memory has been updated, and review your saved memories regularly for anything you don't recognise. That advice still holds, and it applies to any product with memory. The general name for this risk is memory poisoning, and it's the long-term version of what context engineering calls context poisoning.

MemGPT treats memory like a computer's (2023)

Engineers building agents faced a harder version of the same problem. On 12 October 2023, a team at UC Berkeley led by Charles Packer, with Sarah Wooders, Kevin Lin, Vivian Fang, Shishir Patil, Ion Stoica and Joseph Gonzalez, published a paper called "MemGPT: Towards LLMs as Operating Systems". Their idea borrowed from how an ordinary computer works. A computer has a small amount of fast memory and a large, slow disk, and the operating system moves data between them so programs behave as if memory were unlimited. MemGPT gave the model the same arrangement: a limited context window as its fast memory, larger stores outside it, and tools the model could use to save, search and bring back information when it needed to.

They tested it on two jobs that normally break a model: analysing documents far longer than the context window, and chatting with someone across many sessions while remembering what mattered. The project later grew into a company called Letta, and the pattern, a model that manages its own notes, is now standard. On 29 September 2025, for example, Anthropic added a memory tool to its developer platform that lets Claude create, read, update and delete files in a memory folder that persists between sessions, alongside a feature that clears out stale material from the context window. Anthropic reported that, on its own internal test of agents searching for information, using the two together improved performance by 39 percent.

The same idea shows up in consumer products. When Anthropic added chat search to Claude on 11 August 2025, it worked only when you asked it to, searching past conversations on request rather than building a profile. When it launched Claude's memory for Team and Enterprise plans on 11 September 2025, and for Pro and Max subscribers on 23 October 2025, each project got its own separate memory, so a work project and a personal one didn't mix. Products are converging on the lesson from Willison's complaint: memory works best when it has boundaries.

Microsoft Recall and the problem with remembering everything (2024 to 2025)

In May 2024 Microsoft announced Recall, a feature for its new Copilot+ PCs. Recall regularly took snapshots of your screen, analysed them on the computer itself, and let you search your own history in plain language for things you'd seen in apps, websites and documents. Microsoft described it as a kind of photographic memory for your PC.

Security researchers and privacy advocates reacted sharply. Microsoft's own documentation said Recall would not hide passwords or bank details that happened to be on screen, and researchers showed that the saved snapshots could be copied off a machine that malware had already compromised. A searchable record of everything you've seen is a gift to anyone who breaks in. On 7 June 2024 Microsoft announced changes: Recall would be off unless the user chose to turn it on, would require Windows Hello sign-in to view, and its snapshots would be decrypted only when the user proved they were present. On 13 June it went further. Recall would not ship with the Copilot+ PCs launching on 18 June, but would go to its Windows Insider testing programme first. Recall finally began rolling out generally in April 2025, nearly a year later, as an opt-in feature.

Recall isn't a chatbot, but it's the clearest public test of the "remember everything" idea. The lesson carries over directly: the more an assistant remembers, the more valuable and dangerous that store becomes, and the more users need to be the ones who decide what goes in.

An everyday scenario: the memory that outlived the job

This one is illustrative, not a documented event. Imagine Dan, a freelance copywriter. Last spring he spent three weeks writing for a vegan skincare brand, and his chatbot saved several memories along the way: that he "writes for a vegan skincare company", that he prefers "soft, warm language", and that he avoids the word "chemical".

By autumn he's writing for a hardware store. His drafts keep coming back gentle and slightly precious, with an odd tendency to mention natural ingredients. He rewrites his prompts twice before it occurs to him to ask the assistant what it remembers. There's the old job, sitting in every conversation. He deletes the three stale memories, adds one standing instruction about his current client, and the next draft sounds right. Nothing was broken. The memory was just out of date, and it was quietly steering every answer.

Try it yourself

You need a chatbot you already use that has a memory or personalisation feature. Allow twenty minutes. Menu names move around, so if you can't find a setting, search the product's help pages for "memory".

  1. Find the memory settings. At the time of writing, ChatGPT keeps them under Settings, then Personalization, and Claude under Settings, then Memory. Other products use names like "saved info" or "personal context". Note whether memory is on, and whether chat history search is on as a separate switch.
  2. Look at what's saved. Open the list of saved memories and read every item. Then ask the assistant directly, because it may be using more than the list shows:
List everything you know or remember about me from saved memories,
my custom instructions, and past conversations. Group it under
headings. For each item, say where it came from if you can tell.
Don't add anything new or guess. If you don't know, say so.
  1. Check for stale or wrong items. Mark anything that's out of date, wrong, too personal for your liking, or that you don't remember telling it. An item you don't recognise at all deserves a closer look, since memory can be written by content the assistant has read, not only by you.
  2. Delete one item. Remove at least one stale or unwanted memory using the delete control in the settings list. Deleting from the list is more reliable than asking in chat, though some products accept "forget that" as well. Note that deleting a conversation doesn't always delete memories that were saved from it, so check both places.
  3. Add one useful standing preference. Pick something you currently repeat in most chats. Either add it to your custom instructions or ask the assistant to save it:
Please remember this for future chats: I write for a US audience,
I prefer short paragraphs and plain words, and I never want
exclamation marks in anything you draft for me.
  1. Test it in a new chat. Open a fresh conversation, not the same one, and give it a task where the preference should show up:
Draft a three-sentence announcement that my shop is closed
next Monday for a staff training day.

Check whether the preference was followed. If it wasn't, check the memory list to see whether it was actually saved, and consider moving it to custom instructions, which are applied more consistently.

  1. Try a temporary chat. Open a temporary or incognito chat and ask the same question. The preference shouldn't apply there, which is the point: that mode is for conversations you don't want remembered or influenced by memory.
  2. Put a reminder in your diary. Review your memory list every few months, and after any big change in your work. It takes two minutes and prevents a year of quietly skewed answers.

Common mistakes and misunderstandings

Thinking the model itself learns from you. It doesn't. The model is the same for everyone and doesn't change when you talk to it. What remembers you is a separate system of saved notes and searchable history that the product feeds back into each conversation. That's why you can see, edit and delete it, and why it's a feature of the product rather than of the model. Models that truly keep learning after training are a research topic of their own, called continual learning.

Assuming the chatbot remembers everything from a long chat. Within one conversation, older messages can fall out of the context window or get less attention as the chat grows. If something matters, restate it or put it in your standing instructions. For long projects, a fresh chat with a short summary of where you are usually beats one enormous thread.

Treating memory as a place for secrets. Saved memories are stored on the company's servers. Don't store passwords, account numbers, health details or anything about other people that they wouldn't want kept. Whether your conversations and memories can be used to improve the company's models depends on your plan and settings, so check the current terms yourself.

Never looking at what's been saved. Memory builds up quietly. A trip you planned once can turn into "the user lives in Lisbon". A one-off question can turn into a lasting assumption. If answers keep drifting towards an old project, an old client, or an interest you've dropped, the memory list is the first place to look.

Forgetting that memory can be poisoned. If your assistant can read web pages, emails or shared documents, and can also save memories, hidden instructions in that content may be able to write to its memory. Watch for "memory updated" notices you didn't cause, and remove anything you don't recognise. The same caution applies to agents that keep notes: a wrong note gets read back and trusted on every later step.

Wanting maximum personalisation everywhere. More memory isn't always better. For work where you want a neutral answer, such as checking a fact, getting a second opinion, or writing for a different audience, your personal profile can tilt the result towards what it thinks you want to hear. Use a temporary chat, a separate project, or turn memory off for that task.

Confusing memory with project files. Uploaded files in a project are reference material the assistant can look at for that project only. Saved memories follow you across chats. Put client documents in projects, and keep memories for small, durable facts about how you work.

Assuming deleting a chat deletes everything. In some products, a memory saved during a chat stays after the chat is deleted, and the reverse. Check each place separately, and read your product's help page on how deletion works.

Related elements