Continual Learning via Sparse Memory Finetuning

Sparse memory fine-tuning (SMF) addresses catastrophic forgetting, a persistent limitation where large language models lose existing knowledge when learning new information. Traditional methods like full fine-tuning or parameter-efficient approaches like LoRa suffer from either massive interference or limited capacity for new learning. SMF resolves this by replacing standard feedforward networks with specialized, sparse memory layers. By applying a TF-IDF-based ranking metric, the model surgically updates only the memory slots uniquely relevant to new data, preserving general-purpose parameters. Experimental results demonstrate the efficacy of this approach, with SMF limiting knowledge degradation to just 11% compared to an 89% collapse in full fine-tuning. This architecture enables models to continuously integrate new, complex information without compromising foundational reasoning or factual accuracy, paving the way for AI agents that genuinely evolve through long-term experience rather than relying on static, frozen knowledge.