From Cognitive Architectures to Language Agents: A Mechanism-Level Review of Lineage, Convergence, and Migration Gaps

This review compares ten historical cognitive architectures, eight families of language-agent runtimes, and 42 modern systems using seven mechanism dimensions: state, control, transition, persistence, failure, learning, and resource governance. It separates evidence relation (E1-E4) from migration depth (D0-D4). The authors argue that adaptive memory, recovery, team selection, workflow search, skill induction, scheduling, and uncertainty-conditioned action are already operationalized, often through independent convergence rather than documented inheritance. After identifying GraSP as closing one proposed gap, the review isolates five remaining mechanism bundles for falsifiable runtime experiments.