[threecheese]

WRT “Claude learns over time” - this is the biggest gap for me in the current system. As I scale my usage of Claude at work, I observe that it’s quite bad at distinguishing what it should “learn” (memorize) from experimental or just wrong data. It builds and builds on a foundation of sand, making sometimes hidden assumptions and turning them into actionable insight thats just not correct.

It recently wrote an entire dissertation for an epic, assuming it was related to some other project, where it had earlier made the wrong guess about a vendor capability (from their marketing materials, no less), and it all had to be thrown away. I cleared the memory, but it appears to be still pulling from some corporate data source i cant control or locate.

[cyanydeez]

I think memory has very little benefit from non-curated sources. its just a source of halucinatd context and bloated. I think focused work on context removal integrated in the tools would be much better.

Especially in the thinking models; with opencode qnd local LLM. i've got no choice at times but to review the thought traces and if I could go in and correct a few of them before they go do what stupid shit they're hallucinating, I'd reach much better outcomes then fightint with them.