| Already farted around with this in using a free Cloudflare vector database for "memories" with Claude Code for a few months, but LLMs are still a dumpster fire. You can even have it append to a local log file with what it's currently doing, what it's done in the past, conversation history et al. Doesn't really help. If anything, by giving it more cruft to parse through, you're usually making the output even worse. You can witness first-hand how it stores credentials it fetches via the API of a secrets manager for stuff in plaintext too, despite being prompted not to do that. They're a technological dead-end, IMO. Just code the thing yourself, or hire a pro who knows what they're doing. Total waste of time. If you are intending to vibe code a SaaS then you're setting yourself up to get hacked to death along with a world of legal pain like that vuln disclosure post a couple weeks ago here. "Works" is relative, and no, most of the time, it doesn't "work". It may compile, but it doesn't mean it's safe or efficient in any way whatsoever. Claude Code for example does thoroughly bizarre things that no human being would ever do. It cannot reason or deduce, and it is not a debugger. |