... began testing ways to summarise old messages mid-conversation, compressing them just enough to preserve meaning. When I fed those summaries back in, the model continued as if it had never forgotten a thing. That experiment eventually became thredly , a small tool I launched here recently that automates that process, turning long chats into structured memory hand-offs you can reuse anytime. It turns out, memory might be the most underrated part of prompt design. The best prompt ... ... things start drifting. Sometimes I get so mad and say 'stop hallucinating and follow the rules given here' it pisses me off for real (That's why I hardly ask it to do stuffs anymore). Automating the summarization/hand-off like Thredly