Memory Undercutting Model Performance?

I've been pretty disappointed with Fable's performance on my project recently. After looking into Claude's memory, Anthropic seems to neglect memory cleanup and management. If you've used claude on a project for a long period of time, I'd suggest having Codex help you audit it. Here's a summary of what I found today: Recently, I have used Fable a lot to help me with complex prompt engineering tasks. Honestly, it is quite poor at it. This made me wonder if it is also poor at writing to its own memory and succumbs to the same issues. Indeed, it is a mess of rule accretion, misunderstandings, and overstatements. Old instructions remained alongside corrections, narrow feedback was turned into blanket rules, and Claude saved its own explanations as if they were decisions I had approved. I can’t say how much of the performance decline this explains, but my Astra audit found many contradictions and quite a mess. I hope Anthropic would dedicate a little more attention to how persistent memory needs ongoing maintenance. It hasn’t been doing that reliably on my project. Yes, you can disable auto-memory, but there are some aspects of it I would like to use. Makes me curious if Codex handles memory differently.