Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/20/2026

MemT 1.0 Launches: Native Memory Architecture Hits 300B Parameters

3 reports archived0 independent sourcesupdated 10/20/2026, 20:50:34
Synthesis & Latest Updates

Memory Tensor has open-sourced MemT 1.0, the industry's first memory-native foundation model, featuring Flash (288B) and Nano (36B) versions. This release marks a paradigm shift by validating that historical information can be encoded as continuously updated internal parametric memory states at nearly 300B parameters, moving beyond reliance on external retrieval.

Technically, MemT 1.0 integrates a parallel memory computation pathway within the Transformer architecture, enabling the model to directly learn writing, updating, and reading experiences. Built on DeepSeek-V4 (Flash) and Qwen3.6 (Nano) bases, this native architecture offers higher scalability and tighter inference coupling compared to traditional vector database solutions, effectively solving experience retention issues after context window exhaustion.

The release includes full technical reports and code, proposing a Recursive Self-Improvement (RSI) direction where effective methods become the basis for subsequent iterations. This establishes critical infrastructure for autonomous agents with long-term knowledge accumulation, potentially reshaping the developer ecosystem and large model evolution paths.

LATEST/MemT 1.0 has fully open-sourced its technical report and code, demonstrating engineering validation of endogenous memory at ~300B parameters. Current focus is on how its parallel memory pathway empowers Recursive Self-Improvement (RSI) and outperforms traditional RAG in long-horizon tasks.

Hourly heat curve

168 fully observed hours
Now
0.0
Peak
8.715:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 3 reports · Latest first
  2. 机器之心 (微信公众号)T2·85 pts

    MemT 1.0 Launches: First Memory-Native Open Source LLM

    • Released MemT-1.0-Flash (288B/17B active) and MemT-1.0-Nano (36B/3B active), both open-source parametric memory models
    • Key breakthrough: encoding historical interactions into dynamically updated internal memory states to solve experience transfer after context window exhaustion
    • Proposes Recursive Self-Improvement (RSI) direction, aiming to make effective methods the basis for next-round model iterations
  3. 量子位 (微信公众号)T2·78 pts

    MemT 1.0 Released: Native Memory Architecture Scales to 300B Parameters

    • MemT 1.0 features Nano (36B) and Flash (288B) versions, built on Qwen3.6 and DeepSeek-V4 bases respectively.
    • The core innovation is a parallel memory computation pathway alongside attention, enabling models to learn writing, updating, and reading experiences directly.
    • Compared to external memory solutions using vector databases, native memory architectures offer higher scalability ceilings and tighter reasoning coupling.