User as Engram: Local Parametric Edits for LLM Personalization

Personalizing large language models (LLMs) involves two distinct challenges: managing user-specific content and preserving general reasoning skills. While the human brain separates these, with sparse, local memories for episodes and a shared neocortex for skills, current LLM personalization often keeps facts external (e.g., retrieval indices) or merges them globally into weights (e.g., per-user LoRA adapters). The standard per-user LoRA approach, which applies a global weight delta, can inadvertently contaminate text unrelated to the user and has a larger memory footprint. This research introduces "User as Engram," a novel method that stores user-specific content as precise, surgical edits to a hash-keyed memory table within an Engram model. This design ensures that reasoning skill remains in a shared adapter, effectively separating content from skill. This layered architecture not only matches the direct recall performance of per-user LoRA but also achieves 5.6 times higher indirect-reasoning accuracy on average, without degrading the base model's reasoning abilities for any user. The edits are transparent and additive, allowing many users' facts to coexist in a single shared table without interference, a significant advantage over global weight deltas that only accommodate one user at a time.