I keep running into this. I explain my architecture/solution to Cursor in the morning and in the afternoon(if I start a new chat), and I have to explain it all again.
Found something called Recallium (https://recallium.ai) on Reddit last week. Decided to try it out over the weekend.
Curious what others are doing for this problem?
Home grown memory-bank inspired by this: 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.
Oh, nice find! Haven’t seen this one yet.
Looks like it’s Cursor-native, which is probably simpler than my setup. I went with Recallium because I bounce between Cursor and Claude a lot and sometimes VSCode, but if you are mainly in Cursor, this seems like less overhead.
Does it persist well between new chats? That’s what was killing me.
I didnt actually use this one, I started with this one Cursor Memory Bank · GitHub as its only a single rule. But I quickly started figuring out the power of establishing more complicated structures for long term development diary tracking.. Document templates, standardized tools the rules and commands leverage, etc.. all combine to build a robust memory-bank. One of the biggest flaws of the current plan mode in cursor is the fact it doesn’t actively store the plan in the file system as a reference for the user and agents to go back to either to pick up where they left off or to understand what has been done in the past. the TODO list doesn’t provide enough context to be useful to ai or the user half the time. That type of functionality is baked into my memory-bank rule.. consider any of these repos just a starting point that can be customized to your hearts content. 
This is actually quite nice. I am basically just storing raw context in a DB, but you are building a structured system with templates and standards.
The plan mode issue you mentioned is spot on. I have had so many conversations where we plan out a whole feature, then I come back a week later and have no record of what we decided. Just a vague TODO that says implement auth.
So your rules automatically create/update these document templates as you work? And then they are just sitting in your repo, so both you and the AI can reference them?
Yeah, that’s the idea.. I have a planning_templates folder in my repo referenced by my rules as the standard by which new documents should be created. That includes my implementation plans. When I ask the agent to generate a new plan I consider it the core of my larger AI-DLC plan. A starter config if you will. Of course, I’m having to tweak that approach with custom codes being removed, we can establish “rules as agents” based off manual additions to the context, but that doesnt always get the same result IMHO, primarily because important context can be lost when alongside the other context being auto loaded.. But the two together worked pretty well .
The planning_templates + rules combo is basically a documentation framework that Cursor follows automatically.
The context issue you are hitting is interesting
important context can be lost when alongside the other context being auto loaded
This is exactly where I think cross-tool memory systems shine - they use semantic search to retrieve ONLY relevant context instead of loading everything.
Here what might work: Hybrid approach
(your templates): Structure, plans, current work
Recallium: Historical patterns, bug fixes, critical decisions
Only retrieved when semantically relevant Doesn’t pollute the auto-loaded context Works across Claude too if you use it
Your templates give structure. Semantic memory gives smart retrieval.
Right, on with the semantic memory.. but I understand cursor is already adding semantic indexing into the ide, so that should help mane it on the IDE end. But the power of agent custom codes was it felt like they were using system messages for the modes (as they should, but I cant confirm was the case) which is isolated and therefore cannot be polluted by the larger context window.. and guaranteed to run with each call to ensure the correct focus in the task.
Cool find!