Personal Tensor Memory

Large language models (LLMs) excel at general knowledge but struggle when they must remember the preferences, profile facts, and long-term context of a specific user—especially on constrained devices. We introduce Personal Tensor Memory (PTM), a privacy-preserving add-on that assigns every user a fixed-shape matrix, which the frozen backbone can query through one additional attention head. A nightly routine— Hebbian add + decay, norm clipping, slot merge/evict, and occasional orthogonal rotation—re‑organises information inside that matrix without changing its shape or touching billions of backbone weights. On synthetic concept‑drift streams and anonymised personal‑assistant logs, PTM matches kNN‑LM perplexity while needing only 5 % of its context window, and surpasses rank‑8 LoRA under few‑shot data—all using < 8 MB per user and < 1 s daily CPU on a smartphone.