Imagine you're driving a car with a windshield that only shows the present moment — no memory of whether you were accelerating or braking a second ago. That's how most AI-enhanced climate models treat convection: they look at the atmosphere's current snapshot and guess what the unresolved subgrid physics is doing right now. This paper gives the model a rearview mirror. The committed claim: you can learn compact, interpretable prognostic variables — memory states governed by symbolic ODEs — that carry forward subgrid-scale information, and these memory variables substantially improve the temporal structure of parameterized convection, including a realistic diurnal cycle of tropical land precipitation. This is not the first paper to add memory to neural parameterizations, but it is the first to distill that memory into closed-form symbolic equations that can be integrated alongside the resolved dynamics of a climate model. The architecture is a two-stage pipeline. Stage one trains an LSTM autoencoder to compress a window of past high-resolution simulation states into a low-dimensional latent vector (2-8 dimensions). This latent vector feeds into a feedforward neural network that parameterizes subgrid tendencies or surface precipitation. Stage two replaces the autoencoder with a forced multivariate linear ODE whose coefficients are discovered via symbolic regression (PySR). The ODE's state variables become new prognostic variables carried by the coarse model. The key structural choice is linearity of the discovered ODE — they tested nonlinear forms but found linear sufficient for most of the skill gain. The ladder comparison is honest but limited. On Lorenz-96 (a toy chaotic system used as a standard testbed), the memory-augmented parameterization beats the diagnostic (memoryless) baseline on climate statistics and temporal autocorrelation in online (coupled) runs. On the RCEMIP-derived precipitation task, the symbolic-ODE approach recovers most of the autoencoder's offline skill: correlation improvements from ~0.45 (diagnostic) to ~0.55-0.60 (symbolic memory) versus ~0.60-0.65 (full autoencoder). The diurnal cycle of tropical land precipitation shifts from nonexistent in the diagnostic case to qualitatively realistic. No comparison is made to other memory-augmentation strategies in the climate ML literature (e.g., recurrent architectures deployed end-to-end), and no classical convective parameterization with prognostic closures (like prognostic mass-flux schemes) is benchmarked. Integrity is mixed. The Lorenz-96 experiment runs online — the parameterization is coupled back into the resolved model and tested for long-term stability and climate drift, which is the gold standard for this subfield. The precipitation experiment is offline only, meaning the discovered memory variables were never actually integrated inside a GCM. The authors acknowledge this gap explicitly. There is no pre-registration, and code availability is not stated in the abstract. The benchmarks (Lorenz-96, RCEMIP surface precipitation) are community-standard but not adversarial. The milestone that matters for this line of work is online integration of learned prognostic memory variables inside a real Earth system model (e.g., ICON, IFS, or E3SM) with stable multi-year climate runs. This paper demonstrates the concept on Lorenz-96 online and on realistic precipitation offline — the gap between these two is substantial. A plausible next milestone is a single-column model or aquaplanet GCM running with symbolic-ODE memory variables for >1 simulated year without drift, likely achievable within 2-3 years. The obvious experiment the authors did not run is online coupling of the symbolic-ODE precipitation parameterization inside an actual atmospheric model. The honest read is (a) — this requires significant engineering effort and access to a coupled modeling framework, and is almost certainly the next paper. The secondary missing experiment is a head-to-head comparison against prognostic convective closures already used in operational schemes (e.g., the prognostic convective area fraction in ECMWF's IFS). This omission is more telling: it would reveal how much of the 'memory' advantage is already captured by existing physics-based closures.
