Cursor Memory with Project Files and Supermemory MCP
Use Cursor rules, maintained project notes and a scoped MCP connection for persistent context. Verify what is saved, loaded and shared across sessions.
Cursor can use persistent instructions, project files and saved agent sessions. Useful information from an earlier session still needs a path into the current request. When useful context is missing, determine whether it was saved, loaded or retrieved before adding another memory system.
A maintained memory bank is one application pattern. Supermemory MCP is an external context option. Neither automatically turns every past discussion into accurate, authorized context for every future project.
Start with Cursor's native context controls
Cursor's rules documentation describes project rules, user rules, team rules and AGENTS.md. Project rules can be always applied, selected by relevance, associated with files or invoked manually. Review their application mode when deciding which instructions the agent should receive.
Use maintained instructions for stable workflows such as build commands and repository conventions. Use source-linked decision records for why the team selected an architecture. Inspect Cursor's agent documentation for the session behavior of the version you use.
A local file can persist across sessions even if the model never receives it. Conversely, a long-running conversation can contain stale instructions despite retaining history. Saving and loading are separate checks.
Design a small memory bank
You can choose a structure such as:
| Proposed file | Purpose |
|---|---|
projectbrief.md |
Scope and lasting constraints |
activecontext.md |
Current task and unresolved questions |
progress.md |
Completed work and remaining checks |
techstack.md |
Verified tools and relevant versions |
These names are a suggested convention, not a built-in Cursor requirement. A folder named memory-bank does not automatically make Cursor read every file. Reference the relevant files in a rule or request and inspect whether the agent actually uses them.
Keep sources and revisions in the notes. Record “patch prepared,” “merged” and “deployed” as different states. Update a note when the code changes rather than letting a confident old summary override the repository.
Add an external MCP connection where it helps
Cursor's MCP documentation describes server configuration and tool access. For Supermemory's remote server, the setup guide gives this configuration:
{
"mcpServers": {
"supermemory": {
"url": "https://mcp.supermemory.ai/mcp"
}
}
}Use the project or user configuration location appropriate to the intended availability. Complete the OAuth flow, then verify the account and accessible spaces. Enabling a server globally does not mean every repository should share all stored information.
This remote MCP connection is distinct from a client-specific plugin. Installation, capture behavior and authentication can differ. Follow the exact integration's instructions instead of combining a plugin's key setup with the remote server's OAuth flow.
Test capture and recall
The Supermemory MCP reference lists search_memory, get_profile, add_memory, list_spaces and account inspection among its tools.
A planning workflow can retrieve useful context, and an implementation workflow can save a confirmed decision. Configure those steps to fit your workflow and check the tool calls in each session. Cursor's Plan Mode documentation explains how planning fits into the client.
Start by asking which account and space are active. Save one fictional project decision, begin a fresh conversation and ask a question requiring that decision. Inspect the retrieved evidence rather than accepting “I remember” as proof.
Keep cross-project sharing deliberate
Some context can help across repositories, such as a personal explanation preference. A project's credentials, incidents or architectural decisions should not automatically move into another project's answers.
Choose appropriate spaces and verify access. Keep personal context and project-specific records in the intended scopes. Test two repositories with conflicting fictional preferences and confirm which scope supplies each answer.
If importing additional sources, check the current connector list and source-specific sync behavior. Do not assume that every named workspace tool has a connector or that changes propagate instantly.
Budget and maintain the context
Retrieval can reduce the amount of history loaded into a prompt, while adding tool calls and retrieved tokens. Measure the complete task rather than assuming every memory call improves cost or quality. Keep the active evidence set small enough to inspect.
Avoid storing secrets in generated notes. Version-control appropriate team knowledge and keep sensitive local material under the repository's data-handling policy. A .gitignore entry does not remove data already committed.
Test a corrected decision, a moved source file, a removed memory and a request that should not need recall. Recheck the current code before acting on remembered implementation details.
Set up Supermemory MCP for one project and evaluate it alongside maintained Cursor rules. Expand when it helps recover useful context without confusing scope or authority.
Frequently asked questions
Does Cursor automatically read any memory-bank folder?
No. The folder is a convention. Configure or request the relevant reads and verify that the intended context reaches the agent.
Are all Cursor rules always included?
No. Cursor supports different rule application modes, including always applied, relevance-based, file-scoped and manual rules.
Does MCP automatically save every conversation?
MCP exposes the server tools. Actual capture and recall depend on the client, integration and configured workflow. Verify the behavior rather than assuming it.