Externalizing Plasticity: Zero-Update Continual Learning via Symbolic Memories

Foundation models acquire broad competence through large-scale pre-training, yet adapting them to shifting real-world distributions demands either expensive retraining or fragile finetuning—both of which risk catastrophic forgetting of prior knowledge. We ask a different question: must the model change at all? We propose DCASM, a framework for decoupled continual adaptation in which a frozen foundation model is paired with an external DOLCE-guided symbolic memory—a structured “notebook” that records human-ensemble decisions as ontological graphs of endurants, perdurants, and qualities. New skills enter via the notebook, never the model weights; the model learns only to read and execute symbolic graphs. Over time, repeated execution consolidates notebook patterns into model activations, progressively reducing external reliance. On the CLEAR real-world continual learning benchmark, DCASM improves mean accuracy by 8.4 percentage points over vanilla fine-tuning and reduces catastrophic forgetting (BWT: -2.1 vs. -18.4) while performing zero parameter updates to the foundation model at test time