CRISP: Rank-bounded Adaptation for Continual Personalization

Personalized language models must absorb new user information from a growing conversation stream, yet retaining or replaying the full history conflicts with bounded storage and data-retention constraints. Continual adapter training offers a scalable alternative, but its objective differs from task-level continual learning. Conversation batches interleave latent preferences, and later evidence may legitimately revise one preference while unrelated personal information remains stable. We formulate this setting as retentive personalized continual learning and introduce CRISP. CRISP routes each training instance to a small semantic neighborhood of LoRA rank channels and bounds the functional contribution of every channel after optimization. Localized writes reduce unrelated interference, while rank-level projection bounds shared-channel updates without replay, task labels, or growing per-user memory. Under bounded projection, our analysis gives overlap-dependent and stage-count-independent drift guarantees. Across two backbones and interleaved conversation streams, CRISP yields consistent gains in final accuracy and BWT over Vanilla, including a substantial relative final-accuracy improvement on the more densely interleaved stream. These results show that a fixed-size LoRA adapter can balance retention and adaptation without retaining conversation history.