usage of "MEMORY.md" for capturing relevant details

[0ut5ider]

I see that in the latest release the /ce:compound and /ce:compound-refresh use the auto memory feature of claude code to store "relevant" information. PR #311

I want to push back on the current approach for 2 reasons:

1. the auto memory "MEMORY.md" tool is only available to Claude Code, Windsurf and Copilot.
   Yet CE supports many other coding agents (OpenCode, Codex, Factory Droid, Pi, Gemini CLI, Kiro CLI, OpenClaw, and Qwen Code).
   Those agents don't benefit from this memory functionality.

2. The auto memory feature leaves the model to decide what information is "relevant".
   Even if the other coding agents had some sort of memory feature, each coding agent harness will have it's own set of instructions on what the session should store in it's memory file.

I think we need to be more specific about gets recorded in a much more structured approach. After all, the entire COMPOUND ENGINEERING idea is to give the model some solid structured guidance to follow. Why would we leave the memory to be "decided by the model"?

What I think this boils down to it:

• having a structured way to capture the important relevant details (we need to be specific about what is important, and not leave it the model to decide).
  The /ce:work step should be the primary driver of the memory system, and /ce:compound lilkely secondary.
• For file structure of the memories, we should probably have the same name as the plan, but stored in separate folder /docs/memories/ where one file is generated during /ce:work and updated from /ce:compound and /ce:compound-refresh

Anyone else have any thoughts on this?

[tmchow]

PR #311 introduced leverage claude's memory when it exists, we don't use it for storage. The memory is then passed to agents for them to reconcile over that memory in addition to compounded learnings in the repo. We've found the agents are very good at dealing with this.

Our ce:compound (and ce:compound-refresh) skills do the heavy lifting of compounding learnings in your repo and auto classifying and organizing. It also learns to consolidate and when to delete outdated learnings as your codebase evolves.

Given this, can you clarify what the issue is you're reporting?

[andrewlook]

I noticed compound engineering plugin tried to search my entire machine, looking for MEMORY.md - that felt a bit invasive (i got a popup saying "Ghostty.app is trying to access your iPhoto Library" 😳 ). I'd love to be able to disable the search for memory, rather than have the agent try to search every time.

Maybe some kind of ~/.config/compound settings (similar to some suggestions in #338 ) would be helpful.

[tmchow]

I noticed compound engineering plugin tried to search my entire machine, looking for MEMORY.md - that felt a bit invasive (i got a popup saying "Ghostty.app is trying to access your iPhoto Library" 😳 ). I'd love to be able to disable the search for memory, rather than have the agent try to search every time.

Maybe some kind of ~/.config/compound settings (similar to some suggestions in #338 ) would be helpful.

Thanks for the report. Odd it's being so aggressive, will look into it.