Development

An architecture scales to a hundred million tokens with under 9% loss

development2026-03-22

Claim

Chinese lab Evermind AI launched Memory Sparse Attention, showing under 9% degradation scaling from 16,000 to 100 million tokens by decoupling memory from reasoning, while Cursor shipped frontier-level coding at a fraction of the cost.

Reported metric: 100M tokens / <9% loss

Reporting date: 2026-03-22 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to March 22, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to March 22, 2026

Actor

Evidence for

Supersedes

A lab chasing Solomonoff induction gets tenfold data efficiency

Referenced by

Cite or correct this record

Citation formatCitation text

An architecture scales to a hundred million tokens with under 9% loss
Innermost Loop KB — secondary record
Reporting date: 2026-03-22
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-03-22-memory-sparse-attention-to-100m-tokens
Claim: Chinese lab Evermind AI launched Memory Sparse Attention, showing under 9% degradation scaling from 16,000 to 100 million tokens by decoupling memory from reasoning, while Cursor shipped frontier-level coding at a fraction of the cost.
Attributed sources:
- Welcome to March 22, 2026 — https://theinnermostloop.substack.com/p/welcome-to-march-22-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.