Been running some experiments with LLaMA models through OpenRouter, and honestly, the stateless setup is kind of brutal. Having to resend everything with each call makes sense from a routing perspective, but as a dev, it creates a ton of overhead. I’ve already hacked together a small memory layer just to keep context, and it still feels clunky.

Out of curiosity, I tried Backboard.io. It says “waitlist-only,” but I got in fast, so maybe they’re onboarding quietly. What stood out is the stateful sessions, it actually remembers context without me having to do all the duct-tape logic. Makes iterating with local models much smoother since I can focus on the interaction rather than rebuilding memory every time.

Has anyone else here looked into alternatives, or are you just sticking with OpenRouter + your own memory patchwork?