AnswersPublished24 sources
MemT 1.0 Puts Persistent Memory Inside the Model
MemTensor released two memory native models: Flash (288B total, 17B active parameters) and Nano (36B total, 3B active). The models store interaction history in a fixed size internal memory state instead of repeatedly processing the full transcript.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- MemTensor released two memory native models: Flash (288B total, 17B active parameters) and Nano (36B total, 3B active).
- The models store interaction history in a fixed size internal memory state instead of repeatedly processing the full transcript.
- The weights have different licenses: Flash is MIT and Nano is Apache 2.0. The accompanying software has separate terms, so check those before using the code commercially.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Chinese AI company MemTensor announce and open-source on October 22, 2026, with its memory-native MemT 1.0 Flash and Nano models, i. Article summary: MemTensor released the memory-native MemT 1.0 family—Flash and Nano—designed to retain information internally rather than repeatedly process an ever-growing conversation history.[10][11] Two details in the question need . Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
MemTensor announced its MemT 1.0 family—Flash and Nano—as memory-native models designed to keep useful information in a persistent internal state rather than repeatedly process an expanding conversation history. Coverage published on October 22 dates the launch to the evening of October 21.8 The release’s headline efficiency figures come from company-reported tests, so they should be read as benchmark results, not promises of the same gains in every setup.
8
13
What MemTensor released
The two models differ in size and memory configuration. Their total and active parameter counts are not the same as their dedicated trainable memory parameters or dynamic memory state.4
5
| Model | Total parameters | Active parameters | Trainable memory parameters | Dynamic memory state |
|---|---|---|---|---|
| MemT-1.0-Flash | 288B | 17B | 3.61B | 336M |
| MemT-1.0-Nano | 36B | 3B | 451M | 10M |
MemTensor’s model weights are available on Hugging Face. Its official GitHub repository provides implementation materials and links to the technical report, MemT-1.0: Towards Memory-Native Evolution.4
5
20
21
How the models use memory
In a conventional full-context approach, the model processes conversation history as part of the context it receives. MemT’s design instead adds a trainable memory module that updates an internal, fixed-size state after interactions and uses that state to recall information later. The aim is to avoid repeatedly feeding an ever-longer transcript back into the model.7
25
MemTensor says the models were trained on 34 billion tokens in two stages: continued pretraining followed by post-training. The stated goal was to teach the system to write information into memory and answer using what it had stored.8
31
What the reported tests found
In the company-reported long-history tests, Nano’s end-to-end inference was about 1.86× faster than a full-context baseline at 128K tokens of history and 3.99× faster at 256K.8
9 The figures describe specific evaluations; real-world results may vary with workloads and deployment conditions.
13
The reports also compare query time to first token and per-session storage. At 256K history, reported time to first token was roughly 104–112 milliseconds for MemT, compared with about 445 milliseconds for the full-context comparison. The reported persistent memory state was about 40 MiB per session, versus roughly 5.06 GiB for the full-context KV cache—around 130× less storage in that comparison.13
22
25
On 11 memory benchmarks, Flash averaged 50.6 and Nano 47.0, ranking first and second among the parametric-memory methods evaluated. The reported tests did not provide the original history at query time.9
14 The report also describes MemT as competitive with retrieval-augmented approaches, but that finding should not be read as proof that it matches every retrieval system or task.
9
For general capabilities, coverage of the evaluation reports that Flash remained broadly comparable to its pre-memory baseline under distractor-heavy testing, with improvements on some benchmarks.7
25
Check the licenses before using the release
The model weights and software have separate terms. The Hugging Face model cards list MIT for Flash weights and Apache 2.0 for Nano weights.20
21 Materials describing the accompanying inference software list PolyForm Noncommercial 1.0.0; review the relevant repository license and third-party component terms before commercial use.
18
20
21
30
MemT 1.0’s central proposition is that a model can carry forward useful history in a bounded internal state, potentially reducing repeat processing and per-session storage for long conversations. The published tests offer promising measurements for that approach, but they remain company-reported results rather than a guarantee of performance across deployments.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "MemT 1.0 Puts Persistent Memory Inside the Model"?
MemTensor released two memory native models: Flash (288B total, 17B active parameters) and Nano (36B total, 3B active).
What are the key points to validate first?
MemTensor released two memory native models: Flash (288B total, 17B active parameters) and Nano (36B total, 3B active). The models store interaction history in a fixed size internal memory state instead of repeatedly processing the full transcript.
What should I do next in practice?
The weights have different licenses: Flash is MIT and Nano is Apache 2.0. The accompanying software has separate terms, so check those before using the code commercially.