Using cursor rules for agent context, why this is the default flow yet?

Discussions
,
3

Hey, thanks for the detailed post.

Quick clarification on how rules work right now: a rule with globs applies not only when you tag a file yourself, but also when the agent opens or reads a file that matches the glob. That’s part of the standard mechanism. So glob-based rules already support auto-fetch for discovered files. On top of that, rules with a description and alwaysApply: false are pulled in on-demand based on description relevance, which also partly covers the “fetch when needed” case. More details: Rules | Cursor Docs

On the self-managed memory layer side, this idea comes up a lot, and there are a few Feature Requests threads worth checking out and upvoting if it resonates:

If you have specific scenarios where the current glob or description-based fetching doesn’t work well, please share an example. These reports help us understand what to improve next.Hey, nice workflow. A couple quick clarifications on how rules work right now, so you can get more out of them without manual tagging:

In Cursor there are four rule types, see https://cursor.com/docs/rules:

  • Always alwaysApply: true applies to every message
  • Auto Attached globs: in the frontmatter gets pulled in when a file matching the globs is in context, including when the agent opens or reads the file on its own, not only when you manually @ tag it
  • Agent Requested only description: without globs means the agent decides to pull the rule based on the description when it sees it is relevant. This is the closest to what you described, the rule gets fetched only when needed
  • Manual only when you explicitly use @rule-name

So if you want to keep context small but still get the rule when needed, try rewriting some of your rule files as Agent Requested. Remove globs and keep a strong description. Then the model can decide when to pull it in and your context stays cleaner.

Also take a look at Skills Agent Skills | Cursor Docs. They give a similar effect of loadable knowledge modules, but with more explicit structure and for specific tasks.

A managed memory layer without explicit rule files has already been suggested on the forum, for example Your Cursor agent is giving you advice from ages ago — a `cursor-self` MCP would fix this, and the team is aware of the direction. I can’t share a specific timeline for a feature like that, but the feedback is helpful.