What's this about?
Recursive self-improvement stays hard on long tasks because growing histories obscure what the task state actually is and cause the agent to invoke the wrong skills. Recuris proposes a recursive experiential-working-memory architecture for long-horizon harnesses. 21 upvotes.
What exactly goes wrong as history grows?
Everything the agent has done piles into its context. Eventually the current state of the task is buried under a transcript of how it got there, and the agent starts reaching for skills that fit some earlier moment rather than now. Splitting memory into working memory (what is true now) and experiential memory (what I learned) is an attempt to keep the present legible.
What's the community saying?
It pairs with AutoSaddler above and with Prime Agent and Headlong from yesterday's digest — four independent attempts in one week at the same problem, which is a reasonable signal that long-horizon memory is where agent work has converged.