ModelScope releases MemT-1.0: models with built-in persistent memory, 1.86x faster at 128K context
kalyan_kpl · x · 2026-10-20
ModelScope open-weights MemT-1.0, a model family with persistent, evolving memory built into model computation, letting conversations and agents reuse history without reloading the original context.
- Ranks first and second among parametric memory methods across 11 memory benchmarks (vs MemOS, Graphiti, Mem0, etc.)
- Memory lives in the model: dynamic local memory plus trainable hyper memory, no full conversation carried into each inference
- Built for long-running agents: at 128K history, Nano delivers 1.86x end-to-end speedup and 66x lower persistent storage than a full-context KV cache
- Two backbones: Nano on Qwen3.6-35B-A3B (3B active params), Flash on DeepSeek-V4-Flash (17B active params)
More from coding & agent
- How big should your agent's memory file be? One dev caps it at 200 lines with weekly merges — bayvo · 2026-10-20
- Mistral Vibe CLI 3.0 ships with Rust-rewritten TUI: first frame in 72ms, 60fps scrolling — onetwoval · 2026-10-20
- Your agent isn't dumb, your data is lying: audit with 20-30 real questions before scaling — SUPERSAIIYANNN · 2026-10-20
- Building a photo-based bill-splitter: when reasoning models are worth the cost — HowDevelop · 2026-10-20
- Mila's Recursive Self-Aggregation pushes Gemini 3 Flash near top of ARC-AGI-2 — morgymcg · 2026-10-20
- Sparrow turns MRI reports into structured data on infrastructure you control — andrejusb · 2026-10-20