I personally use Mempalace which is perfect to shrink down the size of the memory to be used by Claude Code. It is open source; fully integrated in Claude (plugin instead of MCP); fully working in local (python application and database); fast to be queried and to be installed. My suggestion is to instruct Claude to use it as PRIMARY MEMORY and to save in there everytime you decide that it's time to save or in the end of the session. Also I suggest to save the Mempalace memory database in a separated github repository, so to have it always available for all the devs in the project.

Here is the link: https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf I attach a little summary to the post.

This thread has the purpose to explain the correct ways to prompt Claude and to set it up for managing a whole project from the beginning. I wish to share my workflow and my experience on this so to make community aware of the best practices on this. I couldn't see many good tips around on this topic and also no professional ones. That's why I am happy to start talking about many aspects of a fully controlled project from the very beginning. #specdrivendevelopment #howtostartaproject #claudecode

Hello community! This Skool is intended to stay up-to-date with the Spec Driven Programming tools and methods that will let programmers survive the AI era! Let me introduce my very first contribution. The /effort command let's you decide if the next request will require low, medium, high, max, auto effort from the model to answer: it's a useful way to not waste tokens in tasks that should not require many and to concentrate the max expense of tokens in the important tasks.