User as Engram: Internalizing Per-User Memory as Local Parametric Edits
A research paper and accompanying code that stores a user's facts as a few rows in a content-addressed edit to a language model's parameters, instead of a separate retrieval index or per-user LoRA.
This repository accompanies the paper "User as Engram: Internalizing Per-User Memory as Local Parametric Edits." It splits personal memory into content, a user's specific facts, and reasoning skill, turning facts into answers, storing per-user content as surgical edits to a hash-keyed memory table in an Engram model while a single shared adapter handles reasoning for every user.
Compared with a per-user LoRA on a reference model, the authors report substantially higher indirect-reasoning accuracy at matched direct recall, far less disruption to unrelated text, and a fraction of the storage per user.
The repository is self-contained: it includes the paper's source, the training and evaluation harness, experiment scripts, a per-user-LoRA baseline for comparison, synthetic per-user datasets, and an interactive site built with React and Vite.