AI Agent Memory: What Agents Remember and Where

TL;DR: AI agent memory is whatever an agent can bring into a new task that is not in the current conversation. A model remembers nothing between calls on its own; everything it "knows" about you is text put back into its context window. Short-term memory is the conversation so far. Long-term memory is stored and retrieved: saved notes, searches over past chats, and documents found by retrieval (RAG). The most durable kind is a set of files the agent can search and read, because you can see them, correct them and take them with you.

Disclosure: The Drive AI publishes this guide and makes a file workspace with an agent that keeps notes across conversations and reads the person's files. Vendor features below are described from each vendor's own help pages as of October 2026.

What is AI agent memory?

AI agent memory is any information an agent carries from one step, conversation or task into the next. A language model is stateless: each request starts from nothing but the text it is given. Memory is the system around the model that decides what to store, where to keep it, and what to put back into the prompt next time.

That system has three jobs. Writing: deciding what is worth keeping (a preference, a rule, a fact about a project). Storing: keeping it somewhere outside the model, usually a database or files. Retrieving: finding the right pieces for the current request and placing them in the context. Most differences between products come down to who decides each step: the model, the person, or a fixed rule.

Is a context window the same as memory?

No. The context window is the model's working space: the instructions, the conversation, the tool results and any retrieved text for one request. It is limited in size and cleared when the request ends. Memory is what survives outside it and is put back in.

A large context window delays the problem rather than solving it. A long conversation still fills up, and agent systems then compact it: older turns are summarized or dropped so the recent ones fit. Anything that mattered in the dropped part is gone unless it was written somewhere durable. That is why agents that run long tasks keep notes, plans and results outside the conversation.

What are short-term and long-term memory in AI agents?

Short-term memory is the current conversation or task: what you asked, what the agent did, what the tools returned. It lives in the context window and lasts as long as the session.

Long-term memory persists across sessions. It takes three common forms:

• Saved facts and preferences. Short notes such as "invoices are filed by vendor, then year" or "keep replies short", added to the start of every conversation.
• Past conversations. A search over earlier chats, so the agent can pull in what was discussed last week.
• Documents. A knowledge base or a file store the agent searches when a question needs it.

The first is small and always present. The second and third are large and retrieved on demand, which means the agent has to look and can miss.

What are episodic, semantic and procedural memory?

The three terms come from cognitive psychology and are a useful way to sort what an agent should keep.

The split matters because each type ages differently. Procedures stay true for months. Facts copied out of documents go stale the moment the document changes, so it is usually better for the agent to look them up in the document than to remember a copy.

How does ChatGPT's memory work?

ChatGPT keeps memories across chats and lets you review, delete or turn them off. OpenAI's help pages would not load for us when this was written in October 2026, so for how its settings work today, read OpenAI's Memory FAQ directly rather than a summary of it.

How does Claude's memory work?

Anthropic's help center (October 2026) describes three features that together act as Claude's memory:

• Memory. Claude saves memory as a set of individual topics as you chat, on its own or when you ask. It is on by default for Free, Pro and Max plans; on Team and Enterprise an owner enables it and each member turns it on. You can view, edit and delete topics under Settings → Your Claude, pause memory, or reset it. Incognito chats are not saved to memory.
• Chat search. On paid plans, Claude can search your past chats using retrieval and shows each search as a tool call, with citations back to the original chats.
• Project knowledge. Each project has its own knowledge base of uploaded files and its own memory. A project chat uses that project's memory plus your account memory, but not other projects' memory. On paid plans, when project knowledge approaches the context limit, Claude switches to a retrieval mode that Anthropic says expands capacity by up to 10x.

Project knowledge is the closest of these to the files-as-memory idea: documents the assistant reads when they are relevant, rather than notes it writes about you.

Is RAG a kind of memory?

Yes, it is the retrieval half of long-term memory. Retrieval-augmented generation splits documents into passages, indexes them by meaning, and at question time finds the passages closest to the question and puts them in the context. The agent "remembers" a contract clause because it found it, not because it stored it.

RAG is only as good as what it indexes. A scanned PDF with no text layer, a photo of a receipt or a spreadsheet exported as an image gives retrieval nothing to match, which is why turning documents into searchable text is the first job of intelligent document processing. And retrieval can miss: a question phrased differently from the document may not pull the right passage, so a good agent searches more than once and reads the file before it answers.

Why are files the most durable memory an agent has?

Because they persist on their own, they are the source of truth rather than a copy of it, and the person controls them. A note an assistant writes about you lives inside that assistant's product. A folder of contracts, invoices and meeting notes lives in your workspace, stays current when you edit it, and can be read by whichever agent you use next.

Files as memory have three properties that built-in chat memory lacks:

• You can see all of it. Every file is listed, named and openable. There is no hidden summary of you.
• You can correct it. Edit the document and the next answer uses the new version. Nothing has to be "forgotten" first.
• It is portable. The same workspace can be read by the workspace's own agent, by Claude or ChatGPT over MCP, or by a script. If an outside assistant connects, how MCP authentication works explains what that sign-in lets it read and change.

The trade-off is that files need to be findable. An agent that has to open every file to answer a question is slow; one that can search by meaning, then read the matches, behaves like it remembers.

What does The Drive AI's agent remember?

The agent keeps two kinds of memory, and you control both.

Notes. Short notes about procedures and preferences, not the contents of your documents, that every new request in the workspace starts with. Tell it "Remember that client folders live under /Clients/Client/Matter" and it saves a note; "Forget the note about invoice filing" removes one. It may also write a note on its own when it notices a way of working you repeat. Notes are either yours (only your requests, only you see them) or the workspace's (everyone's requests). Only owners and admins can save workspace notes, and the agent asks first with a card warning that "Every member's assistant will follow this from their next conversation on", so text inside a document cannot quietly rewrite how everyone files. Each note holds up to 4,000 characters, and all of them are listed, editable and deletable in Settings → Agent → What the agent remembers. Notes belong to one workspace and do not follow you into another. Asking it to remember or forget does not count as a task.

Files. Everything else the agent knows comes from the workspace itself. It searches files by what they are about and reads the ones it needs, so a question about a lease is answered from the lease, as it is today. Owners and admins can add standing workspace rules and a persona in the same settings tab.

Your conversations are private and stay in History for you to reopen. What carries into a new conversation is the notes, the workspace rules and the files.

Frequently Asked Questions

Do AI agents remember previous conversations?

Only if the product stores something and puts it back. Some assistants save notes about you, some search your past chats, and some do both. Without one of those, each conversation starts from nothing. Check the product's memory settings to see what it keeps, and whether you can view and delete it.

What is the difference between AI memory and a context window?

The context window is the text a model can see during one request: instructions, conversation, tool results. It is cleared afterwards. Memory is stored outside the model and placed back into the context window in later requests. A bigger window holds more at once but does not keep anything after the session ends.

Is RAG the same as long-term memory?

RAG is one way to implement it. Retrieval finds relevant passages in a store of documents and adds them to the prompt, so the agent can use information it never saw before the question. Long-term memory also includes saved notes and preferences, which are small enough to include in every request without retrieval.

Can I see and delete what an AI agent remembers?

In most current products, yes. ChatGPT and Claude both let you delete saved memories and turn memory off, and Claude lets you edit each memory topic. The Drive AI lists every note under Settings → Agent → What the agent remembers, where you can edit or delete each one. Files are the easiest memory to inspect, because they are simply your files.

Should an agent remember facts from my documents?

Usually not. Facts copied into memory go stale when the document changes, and the agent can look them up again. Keep memory for procedures and preferences, such as naming rules or where client folders live, and let the agent search and read the documents for facts.

Agent memory works best split in two: a few short notes about how you work, and a workspace of files the agent reads for everything else. To try both, upload a folder to The Drive AI, tell the agent one rule you want it to keep, and open Settings → Agent to see the note it saved. For filing at scale, the AI file organizer sorts a whole folder by those rules and shows the plan before anything moves.