I started AI-assisted coding quite a while ago with "query for code to copy and paste" approach which was slow but it dramatically shifts when the LLMs are used as agents that are just AI that have access to certain things like your project's source codes, internet and some other technical docs that refines them. You can simply instruct it to change snippet of codes by mentioning them with their actions inside the chat that feeds the agent, this is done in tools like cursor, antigravity, llmanywhere. an instruction could be limited to CRUD instructions, CRUD standing for Create, Read, Update, and Delete. an update instruction looks like "change the code that does this to do that" or more precise one "change the timeout of the request to ycombinator.com to 10". having a good memory definitely helps here but forgetting isn't the end of the development or necessity to start reading the source codes yourself to know where an instruction should target but you can ask the project's interconnected source codes (i put interconnected because it generates lots of source codes for some cases like test cases that aren't used in production but are part of the project in my experience with cursor for example) goal summary if you've forgotten the big picture of the project because you came back from a break or something. I used AI agent for my last langgraph solo project only which had python and go languages, git and cursor so take my advice with a grain of salt :)