Cursor memory: what Cursor keeps, what changed, and four places a memory can live

Cursor remembers through rules that you write, and its own Memories feature no longer has a settings page. That leaves four places a memory for Cursor can live. Here is our own proposal for choosing between them, the list test: before you trust a memory, ask whether you can list everything in it, where it is stored, and which other tools can read it. A memory you cannot list is one you cannot correct.

The four places, side by side

What changed in Cursor's own Memories

Cursor shipped Memories in 2025 as rules that it wrote for you from your chats. Three dated facts describe where that stands now.

• On November 25, 2025 a user on Cursor's forum reported that Memories had disappeared from settings in version 2.1.32 and that going back to 2.0.77 restored them. A reply the same day, from Dean Rie, said the feature had been removed on purpose starting with 2.1, and pointed to an export command that saves existing memories as a rule file.
• On January 10, 2026, in a second thread, the same person replied that the feature had not been removed: the settings page was gone, and memories were to be added or deleted by asking the agent in chat. The forum pages do not show his role at Cursor.
• On November 11, 2026 the address that used to hold Cursor's Memories documentation showed the Rules page, and that page does not mention memories. It lists four kinds of rule: project rules, user rules, team rules and AGENTS.md.

Those three do not fully agree, so check your own version. What they do agree on is the part that matters for the list test: there is no settings page where you can read what Cursor has remembered.

Why list first

A memory changes what the agent does in every later chat. If it holds something wrong, such as an old convention or a decision you reversed, the agent keeps acting on it until someone removes it. You can only remove what you can see. That is why the first question is whether you can list the whole memory, before how clever its search is.

The other two questions follow from the same idea. Where it is stored tells you who else can read it and whether it leaves with you when you change tools. Which other tools read it tells you whether you are building one memory or one per editor.

What we counted on November 11, 2026

We read every server in the official MCP Registry through its public list, asking for the latest version of each, and looked for the whole word "memory" or "memories" in each server's name, title or description.

Two limits. A word match counts servers that mention memory in passing, and misses memory servers that never use the word. And a package you run on your own computer can still send what it stores to a company's servers, so "package" does not mean "kept on your computer". The count shows how much choice there is, which is the reason to have a test. ModelBrain is one of the 70 offered both ways.

Place 1: rules

Rules are instructions Cursor loads into the agent's context. Project rules are files under .cursor/rules and travel with the repository. User rules apply across all your projects and are set in Cursor's settings. Team rules are managed from Cursor's dashboard on Team and Enterprise plans. An AGENTS.md file in the project is plain Markdown.

Rules pass the list test easily, because you wrote them and can read them. Their limit is that they are standing instructions, not a record. Nothing adds to them unless you do.

Place 2: Cursor's Memories

Whatever state the feature is in for your version, it fails the first question today: you cannot open a page and read the list. If you used Memories before version 2.1, the forum reply above describes an export command that saves them as a rule file, which turns them into place 1.

Place 3: a memory bank

A memory bank is a pattern from Cursor's community, not a Cursor feature. You keep a few Markdown files in the repository, for example the project brief, the current state and the decisions made, and a rule tells the agent to read them at the start of a task and update them at the end.

It passes all three questions: the files are listable, stored in the repository, and readable by any tool that opens the folder. The costs are that every file the agent reads takes room in its context, the files grow, and anything in them is committed with the code unless you exclude it.

Place 4: an MCP memory server

Cursor is an MCP client, so a memory server gives the agent tools to save and search memories outside the repository. With 1,340 to choose from, apply the test to each one. Does it have a tool or a screen that lists every memory, and can you delete one? Does it keep the data on your computer or on its own servers? Does it work with the other assistants you use?

ModelBrain's answers: the list_memories tool lists them and forget removes one; the memory is stored on your own computer; and the same memory serves Cursor, Claude, ChatGPT and any other MCP client you connect. Setup writes Cursor's MCP entry for you, as described in the Cursor setup guide. The limits: it needs the ModelBrain app installed, and it holds facts and decisions, not editor rules, so it adds to place 1 and does not replace it.

Which one to pick

• Standing instructions for a repository, such as style, commands and what not to touch: project rules or AGENTS.md. Use AGENTS.md if other agents work in the same repository. See our AGENTS.md article.
• Your own habits across every project: user rules.
• Project context that should travel with the code: a memory bank.
• Facts and decisions you want every assistant to know, kept out of the repository: an MCP memory server that passes the list test.

What this does not cover

This is about what Cursor carries between chats, not how much RAM the app uses. We did not install or test the 1,340 servers; the count is from their registry entries. We have not confirmed the state of Memories in the newest Cursor build ourselves; that section reports Cursor's forum and documentation as we read them on November 11, 2026.

Sources

• Rules, Cursor documentation
• Are my memories gone?, Cursor community forum, November 25, 2025
• Memory tool disappeared, Cursor community forum, January 9, 2026
• Server list, official MCP Registry
• Give Cursor a persistent memory with ModelBrain, modelbrain.net

Common questions

Does Cursor have memory? Cursor carries instructions between chats through rules: project rules, user rules, team rules and AGENTS.md files. It also had a feature called Memories that wrote rules from your chats. According to replies on Cursor's forum, the settings page for Memories was removed in version 2.1, and Cursor's current rules documentation does not mention memories.

What happened to Cursor Memories? A user reported on November 25, 2025 that Memories had gone from settings in version 2.1.32. A forum reply that day said the feature was removed on purpose from 2.1 and pointed to an export command. A reply on January 10, 2026 said the feature still existed and was managed by asking in chat. Check your own version before relying on either.

What is a Cursor memory bank? A pattern from Cursor's community, not a Cursor feature: Markdown files in the repository that hold project context, plus a rule telling the agent to read them at the start of a task and update them at the end. You can read and edit every line because they are ordinary files.

Can Cursor use an MCP memory server? Yes. Cursor is an MCP client, so a memory server added to its MCP settings gives the agent tools to save and search memories. On November 11, 2026 the official MCP Registry listed 1,340 servers that mention memory, and 45 of those name Cursor.

Is a memory shared between Cursor and Claude? Rules files in Cursor's own format are for Cursor. An AGENTS.md file is plain Markdown that other coding agents also read. A memory server is shared by every MCP client you connect to it, which is how one memory can serve Cursor, Claude and ChatGPT.

Last checked November 11, 2026. Cursor changes often, so confirm a detail in your own version before relying on it.