Persistent, intelligent project memory

Feature Requests
1

.cursorrules is a stopgap.

What we really need is for Cursor to truly remember interactions with the user and what the project needs and auto-update this memory as the user interacts with Cursor and the project.

Here’s an example of a great stab at this from the Cline community.

Here’s a far more advanced example of the same idea.

This would be HUGE.

It would improve Cursor significantly by

  • helping reduce prompts like “No, I told you to…”
  • remembering architectural details/decisions and/or implementation details about the project consistently
  • reducing or completely eradicating model “stray” where it goes of and does something random and unrelated to the user’s intent
  • countless other ways

OpenAI has solved this rather elegantly.

10 Likes
2

Tried the Cline one, works like a charm. Might try it tomorrow on something bigger.

3

This is wonderful, thank you … weirdly - it said “Welcome back!”

anyway, i (re?) signed-up…

As I’ve been vocal about I (and pretty much everyone it appears) are all reaching for the same, next, goal-post in the evolution of AI’s Entanglement into Human Civilizations Technological Mycelium that intertwines us all…

To harness and keenly wield Agentic Computing, and as its tagged on Mem0: Companion Computing…

–

Lokoing just at the vid intro landingpage for Mem0: immediately, the gleaning of the data from the convo its monitoring is interesting – but what I want is to do somethign a little more interesting than that - let me know if it will help enable the following:

I want Mem0 as an agent in my swarm of agents which will watch all my other agents and my projects and my various workspace rules-sets and such - and it will be the companion bot that then harvests all my meta data into my project orchistration control plane – and so it keeps a running context of all my tinkerings, in a github repo – basically every action is repo’d and persisted to postgres for meta-layers of iterating through all my projects from a singular WorkBench.

Thats the direction I am looking for… Does this align with everyone as much as it appears to?

4

Thanks for the request. This is definitely something we’re interested in and seems like a gap we have yet to fill.

We are always experimenting with stuff behind the scenes and if we find something that works well we will be sure to work on it and ship it to you when we can.

However, we don’t have anything ready to announce right now!

2 Likes
5

I came across this and wanted to share: LLM Memory: Integration of Cognitive Architectures with AI

6

Had a workaround with SCRATCHPAD.md file for this purpose, but LLMs does not always pay attention to it in the chained requests, unless you have enforced workflow-level rules to do this explicitly upon each interaction.

Another example - Cursor’s Memory Bank

Recently this can now be implemented with MCP memory (knowledge graph),
however that consumes extra LLM call and does not do related context recall automatically (auto-RAG style)

7

This also looks promising btw: Cursor IDE: Adding Memory With Graphiti MCP 🤖⚡️

8

Any updates on this? OP described the necessary function beautifully – until we can truly have that functionality in, why not something like Windsurf has? “Memories can be manually or automatically defined, and are persisted as context to better align the Cascade’s outputs with the user’s preferences.”

Seems a bit more simple than OP’s more broad suggestion, more

But Mr. Dan, just SOME automatic/manual memory and context function built in so I can ditch the halfbaked MCPs would be amazing. This is one of the singular things that have made me think about leaving Cursor.

Edit: No shade to the MCP developers, the ones that work well work well! We just need an actual implementation.

9

This is something we are looking at internally, but nothing formal to announce yet unfortunately.