Survey of Cognitive Architectures and Language Agents Reveals Mechanism-Level Gaps

So, there's this arXiv preprint, numbered 2607.23942, that dives into ten historical cognitive architectures alongside eight types of language-agent runtimes and forty-two current systems that focus on mechanisms. The authors break down each mechanism by looking at aspects like state, control, transition, persistence, failure, learning, and resource management, while also distinguishing between evidence relations (E1-E4) and migration depth (D0-D4). They found that today’s agents are successfully using things like adaptive memory and failure recovery, mainly through independent convergence instead of inherited traits. The review points out that there's room for better integration of these mechanisms and aims to clear up confusion around terms like memory and planning.