Neural memory as an unbounded 2D lattice: writes and reads through a tiny 3x3 bump
CFGeek · x · 2026-10-03
A team led by Christos Tzamos unveiled a novel neural memory architecture, detailed in a blog post and paper.
Memory is modeled as an unbounded 2D lattice of recurrent cells. The network learns to map each key/query to a point on the lattice, and performs writes and reads only through a tiny 3×3 bump around that point. Commenters call it a remarkably clever take on memory — a core bottleneck for long-context LLMs and persistent agents.
More from Research
- ezyang's DeepSeek-V3 series part 3: roofline analysis for training DSv3 on Hopper — ezyang · 2026-10-03
- 36B Open-Source Robot Model Isaac 0.5 Runs Real-Time Under 50ms Without Local GPU — AkshatS07 · 2026-10-03
- Harmfulness eval pipeline: probes on BeaverTails, training on JBB, OOD eval on ClearHarm — maksym_andr · 2026-10-03
- A robot demonstration can lose its most useful half-second: tracking gaps sabotage manipulation learning — Klutzy_Cap8492 · 2026-10-03
- Hinton: leading researchers now think an intelligence explosion may happen soon — geoffreyhinton · 2026-10-03
- Spotlight architecture scrutiny: attention-style memory may carry a large constant factor — teortaxesTex · 2026-10-03