Losing context across files in Cursor, how are you handling this?

Discussions
1

I’ve been using Cursor a lot for feature work, and it’s great when I’m focused in one area. Where I keep hitting friction is when a task spans multiple files or I come back to it later earlier architectural decisions don’t always carry forward, and I end up re-explaining or correcting things that were already agreed on.

I’ve tried leaving better comments and keeping chats focused, but once the workflow becomes write → refactor → tests → cleanup, it starts to feel fragile. Debugging turns into scrolling through old messages instead of reasoning about the code.

Recently I started testing a small tool called Zenflow to keep a shared spec/state that the agent can reference across steps. I’m using it alongside Cursor (not replacing it), and it’s helped a bit with longer-lived tasks, though it’s still early.

Curious how other Cursor users handle this. Do you keep sessions short, rely on external docs, or have some workflow that works well for bigger changes?

2

I use a custom built, rules-based, project level, memory-bank to keep track of plans, implementations, architectural decisions, etc.

Mine was inspired by GitHub - vanzan01/cursor-memory-bank: A modular, documentation-driven framework using Cursor custom modes (VAN, PLAN, CREATIVE, IMPLEMENT) to provide persistent memory and guide AI through a structured development workflow with visual process maps. and this one Cursor Memory Bank · GitHub . But pretty quickly mine changed to follow my flow.

3

I talked with a guy on X about it who recommended just chucking it all in AGENTS.MD which I do a fair bit, to me it shouldnt be human readable and just a bucket of context.
I am not convinced AGENTS.MD works without it being tagged its a bit of a black box to me - for example I ask it to never useMemo and useCallback but it loves to throw it in anyway.

I have another one called “new-feature.md” which goes through every full-stack thing that neds to be done to add a new feature which I find helpful too.

4

Yeah.. not sold on AGENTS.MD as being consistently implemented or read. also long term I have concerns about them getting too big to be useful (depending on the solution)