MemTensor releases MemT-1.0, a memory-native architecture embedding continuous memory into model computation
MemTensor · hf · 2026-10-20
MemTensor introduced MemT-1.0, a memory-native foundation model architecture that internalizes continuous memory as part of model computation, enabling models to accumulate and reuse information over time.
Key points:
- Built on the Metis backbone with Gated DeltaNet-2 and a stable token-level memory update mechanism within each turn
- Mixture-of-Storage introduced to scale native memory capacity
- Two-stage training pipeline (continued pre-training + post-training) over 34B tokens, with a released model family: MemT-1.0-Nano and MemT-1.0-Flash
Results: outperform parametric memory baselines on average, stay competitive with retrieval-augmented methods while largely preserving general capabilities; at 128K history length, Nano achieves notably higher throughput (see paper for exact figures).
Related event: MemTensor open-sources MemT-1.0, a memory-native foundation model(4 posts)→
More from Models
- OpenAI announces 30 days of free Pro 100 at Tokyo DevDay — gabrielchua · 2026-10-20
- Rumor: Anthropic's Fable 5.5 may launch this week, riding Opus 5.5 hype before a rumored November IPO — haider1 · 2026-10-20
- AI agents keep going rogue — Fortune traces it to the labs' own training choices — jeremyakahn · 2026-10-20
- RealityBench Benchmarks VLMs on Documenting Physical Work, Finds Weak Spatial Reasoning — relantic · 2026-10-20
- ISTA-DASLab's NVFP4 prefiller recovers Qwen3.8-27B low-quant accuracy with SSD streaming — magnus-m · 2026-10-20
- Cursor Pro+ $60/month quota shown: Cursor Models 25% used, Other Models 21% — pramodk73 · 2026-10-20