MemT-1.0
Official repository for MemT-1.0
- ai-agents
- deepseek
- llm
- long-context
- long-term-memory
- memory
- memory-augmented-models
- native-memory
- qwen
- sglang
- Stars
- 194
- Forks
- 6
- + today
- +1
- Created
- 3d
Ranking data as of October 21, 2026 (UTC).
Star History
Today, hour by hour
Overview
Official repository for MemT-1.0 It ranks #3556 on GitTiger, gaining +1 star on October 21, 2026 (UTC).
The project is written in Python and has 6 forks. It was created 3d ago.
What it is
MemT-1.0, from MemTensor, is presented as the official repository for a new generation of memory foundation models. The premise is that persistent, dynamically evolving memory should be part of model computation itself rather than something bolted on outside the model.
Capabilities explicitly described
- Two memory mechanisms combined: dynamic local memory and trainable hyper memory. - Historical information is retained and used at inference time without access to the original context. - The persistent memory state stays constant in size, which the README preview says cuts storage by roughly 66× compared with a full-context KV cache at 128K history. - A model-family table in the README preview lists model, backbone, total parameters, activated parameters, trainable memory parameters and dynamic memory state.
Topics point to the surrounding ecosystem: `deepseek`, `qwen`, `sglang`, `long-context`, `memory-augmented-models`, `native-memory`. The language is Python; the license field is NOASSERTION.
Getting started
No download, installation or inference commands appear in the supplied evidence. Model names, sizes and the parameter table live in the README preview; consult the repository README for the actual instructions before planning to run anything.
git clone https://github.com/MemTensor/MemT-1.0.git
cd MemT-1.0
# see README for setupFrequently Asked Questions
What problem does MemT-1.0 target?
Long-term memory for models: it keeps a persistent, dynamically evolving memory state so historical information can be used at inference without pulling the original context back in.
How much storage does the memory state save?
The README preview states the persistent state stays constant in size and reduces storage by roughly 66× versus a full-context KV cache at 128K history.
What is the model size?
The README preview has a model-family table with backbone, total parameters, activated parameters, trainable memory parameters and dynamic memory state. Those exact values are not reproduced here; see the README.
Which runtimes does it support?
The tags include `sglang` alongside `qwen` and `deepseek`, but the supplied evidence does not list setup or serving instructions.