MemT-1.0-Flash

Use it for text generation that draws on persistent conversational memory, or for memory-guided coding tasks.

Availability

Model weights are available on Hugging Face.

Pricing model

Not established

Deployment

The quick start loads the model locally with Transformers.

Privacy

The memory-enabled example clears plaintext conversation context after saving information to latent local memory.

Licence

MIT License for MemT-1.0-Flash weights; backbone and third-party components retain their original licenses.

• MemT-1.0-Flash integrates dynamic local memory and trainable hyper memory.
• A MemT-1.0-Flash model checkpoint is linked on Hugging Face.
• The inference guide provides setup instructions for MemT-1.0-Flash. ^[1]

  For deployment details and MemT-1.0-Flash setup, see the inference guide in [English](inference/README.md) or [简体中文](inference/README_zh.md).

• MemT-1.0-Flash weights are released on Hugging Face.
• The model integrates dynamic local memory and trainable hyper memory. ^[2]

  MemT-1.0 integrates dynamic local memory and trainable hyper memory to retain and leverage historical information without accessing original context at inference.

• Its persistent state remains constant in size, with approximately 66× less storage than full-context KV cache at 128K history length.
• The model supports saving and resuming local memory without loading the plaintext conversation. ^[2]

  # Restore local memory without loading the plaintext conversation.

• Across 11 memory benchmarks, it ranks first among the evaluated parametric-memory methods. ^{[1] [2]}

  Across 11 memory benchmarks, MemT-1.0-Flash and MemT-1.0-Nano rank first and second among the **evaluated parametric-memory methods**, respectively.

• The MemT-1.0-Flash model weights are released under the MIT License. ^[2]

  The MemT-1.0-Flash model weights are released under the