Self-Growing On-Device AI Model Expanding Parametric Memory Without Rewriting Factual Knowledge

Personal AI should adapt to how its user works without forgetting what it already knows. Yet full fine-tuning is too costly for mobile devices, cloud adaptation exposes private data, and even lightweight post-training can distort factual knowledge or overwrite earlier preferences. We present Synapse, an on-device personalization system that stores recurring user conventions in scope-gated low-rank synapses around a frozen backbone. A convention that cannot be applied from context is trialed against existing synapses under factual and persistence checks. Synapse reuses an admitted synapse, grows one when all eligible synapses fail admission and the storage budget permits, and otherwise leaves the convention in retrieval. Across nine sequential writes with evolving and conflicting conventions, a capacity-matched continual adapter improves from $23\pm8\\%$ to $51\pm8\\%$ accuracy on earlier, non-revoked conventions when given persistence replay. With the same 27.0-MB state, factual fence, replay-token budget, and cumulative optimization budget, Synapse reaches $79\pm6\\%$, a further 28 percentage points. Allocating a synapse per write reaches $82\pm5\\%$ at $3.0\times$ the storage. TriviaQA/GSM8K flip rates are $1.7/2.5\\%$ for Synapse and $1.9/2.7\\%$ for the replay-matched continual adapter; removing Synapse's fence raises its flips to $5.6/8.3\\%$. Synapse training runs locally during idle charging on an iPhone 13 Pro; the complete protected write, including fence and admission checks, is measured on an Apple M4.