MemT-1.0 gives language models a built-in memory

What this is about

The research team behind MemT-1.0 introduced two open language models on October 19, 2026 that do not merely carry memories as old chat text. Instead, they write information into a compact model state during a conversation. The approach is intended to help long interactions and AI agents reuse experience across multiple work steps.

For MemT-1.0-Nano with a 128,000-token history, the technical report claims roughly 1.86 times faster end-to-end processing and 66 times less persistent storage than a full KV cache. These are measurements from the authors, not independently confirmed production results.

What MemT-1.0 actually does

MemT-1.0 adds two memory components to a backbone model. A local memory changes during use and retains compressed information from earlier segments. A trainable Hyper Memory Block uses read, write, erase and forget mechanisms to control what enters this state. The authors call the scalable combination of these stores Mixture of Storage.

The team released MemT-1.0-Nano with 36 billion total parameters and 3 billion active parameters, plus MemT-1.0-Flash with 288 billion total and 17 billion active. Nano uses Qwen3.6-35B-A3B as its backbone, while Flash uses DeepSeek-V4-Flash. According to the model card, Nano's weights are available under Apache 2.0. A session can be saved as a state file and resumed later without reloading the original conversation.

Why it matters

Many assistants currently retain context by sending earlier messages to the model again or retrieving material through an external search system. Both approaches consume storage, computation and, in long histories, attention within the context window. A compact state could lower the cost of services where people work on projects for weeks or agents are expected to learn from earlier attempts.

Across eleven memory benchmarks, the two MemT models ranked ahead of the parametric-memory systems tested by the team. At the same time, the official model card explicitly says textual-memory systems still lead on long histories that require detailed evidence retrieval. That distinction matters: compression may preserve useful patterns, but it does not replace an auditable archive.

In plain language

Imagine a cook who does not keep every order ticket after each evening, but updates a small notebook: less salt for table twelve, start the soup earlier, remember a regular customer's allergy. The notebook takes far less space than every individual ticket. But if someone later asks for the exact wording of an order, the original ticket is more reliable.

MemT-1.0 tries to do the same with model histories. It compresses experience into an evolving state instead of carrying the entire record every time.

A practical example

A coding assistant accompanies a team through 100 tasks, each with 2,000 tokens of history. A conventional approach may need to store or search large parts of those 200,000 tokens when answering later questions. MemT-1.0 could instead retain rules such as “run tests first” or “this project uses Python 3.11” in its memory state.

On the next task, the assistant can use that compressed experience without reading every old conversation. The paper reports improvements over the respective backbone models on experience-guided coding tasks. That does not establish that every real codebase will see the same gain.

Scope and limits

• The benchmarks and efficiency figures come from the development team. Independent replications and measurements under production load are not yet available.
• Compressed memory can lose details or preserve a wrong inference. Contracts, diagnoses and auditable decisions still need traceable original sources.
• The smaller model has 36 billion parameters and its instructions call for Python 3.11, CUDA 13 and a suitable GPU. It is not a lightweight model for ordinary laptops.
• The model card itself says textual-memory systems still lead on long histories requiring precise evidence retrieval.

SEO & GEO keywords

MemT-1.0, MemTensor, memory-native AI, AI memory, language model, KV cache, Mixture of Storage, Gated DeltaNet-2, Qwen3.6, DeepSeek-V4-Flash, long context, coding agent