MemT-1.0
Memory-native LLMs (36B and 288B MoE) that keep history in learned memory, claiming 66x less storage than a KV cache at 128K tokens.
Overview
MemT-1.0 is a pair of memory-native language models from the MemTensor group, described in an arXiv paper dated October 19-20, 2026. MemT-1.0-Nano (36B total, 3B active, built on Qwen3.6-35B-A3B) and MemT-1.0-Flash (288B total, 17B active, built on DeepSeek-V4-Flash) add a Gated DeltaNet-2 layer with token-level memory updates and a Mixture-of-Storage mechanism, after 34B tokens of continued pretraining and post-training. The authors report a 66x reduction in persistent storage and 1.86x end-to-end speedup versus a full-context KV cache at 128K tokens, and averages of 50.6 (Flash) and 47.0 (Nano) across eleven memory benchmarks. Code is under PolyForm Noncommercial and the paper under CC-BY-NC-SA, so commercial use needs a separate license; results are self-reported.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Nano: 36B total / 3B active, 10M dynamic memory; Flash: 288B / 17B, 336M dynamic memory
- Gated DeltaNet-2 layer with token-level memory updates
- Mixture-of-Storage mechanism to scale native memory capacity
- 66x less persistent storage and 1.86x faster inference than full-context KV at 128K
- TTFT speedup of 2.44x at 128K and 4.19x at 256K
- Code under PolyForm Noncommercial; paper CC-BY-NC-SA
Use cases
- Long-running agents that must retain history without growing KV caches
- Research on memory-native architectures
- Coding agents that accumulate experience (+7.6 on SWE-bench Verified over backbone, self-reported)
Pricing
Free for non-commercial use; commercial use requires a separate license from the authors.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 19 Oct 2026
- Record updated
- 28 Oct 2026
- Record last checked
- 28 Oct 2026