MemTensor's MemT-1.0 report: 66x less storage than full-context KV cache at 128K tokens
jiqizhixin · x · 2026-10-29
MemTensor released the MemT-1.0 technical report, making memory part of a foundation model's computation. The Nano and Flash variants pair standard attention with a parallel memory pathway: learned gates write conversation info into a persistent matrix state and read it back on demand, stabilized by Gated DeltaNet-2 and scaled via Mixture-of-Storage, with no backbone fine-tuning needed. At 128K tokens of history, Nano reports 66x less persistent storage than a full-context KV cache and 1.86x faster end-to-end inference in a two-GPU single-request setup. Flash and Nano rank first and second across 11 memory benchmarks among parametric-memory methods.
More from Infra
- Communication Is the Bane of Inference and Training: One Beefy GPU Beats Multi-Node — TheZachMueller · 2026-10-29
- Tailslayer: open-source C++ library kills DRAM refresh stalls, cutting p99.9 latency up to 18.5x — thisdudelikesAI · 2026-10-29
- Analyst: skip Qualcomm Dragonwing—it's branding, not a real platform — DavidLinthicum · 2026-10-29
- Europe's data center boom sparks a fight over who gets grid power first — kindermaxi123 · 2026-10-29
- Reddit asks: what have you actually built with local LLMs? — Loose_Doubt367 · 2026-10-29
- Minigraf: SQLite-style single-file bi-temporal graph database for AI agent memory — JeremyCMorgan · 2026-10-29