Experiment: a learned memory table transfers across model sizes with a small adapter

sn2006gy · reddit · 2026-09-27

The author tested an 'Engram' — an external memory table of associations a model can look up instead of storing facts in weights. A table learned by three small models on synthetic key-to-symbol tasks was attached to two differently sized models of the same family. Mixed results: the exported table was byte-identical across sizes; used directly it matched the no-memory baseline (3.125% accuracy); after training a small adapter per recipient, accuracy reached 71.7% on one held-out wording but stayed at 3.125% on another, with one of six adapter runs missing the pre-set 50% threshold. Evidence that memory tables can move between model sizes and become useful with local tuning — but far from general-purpose transfer. Code and protocol are open-sourced on GitHub.

Original post →

More from Research

Research channel →