MemT-1.0-Nano GGUF

GGUF conversion of MemTensor/MemT-1.0-Nano (Qwen3.6-35B-A3B backbone + MemT native memory).

The K-quants are made from a BF16 conversion without imatrix. In all files the MemT memory weights (memt_*) are kept at Q8_0, because the memory state accumulates their errors over every committed token.

Requirements

These files contain the MemT memory weights (blk.*.memt_*, keys qwen35moe.memt.*). Upstream llama.cpp does not support them yet and will refuse to load them.

Use the memt-native-memory branch of Brishen/llama.cpp. It is current upstream llama.cpp plus one commit that adds the MemT memory (CPU and CUDA), the converter and the llama-server API:

git clone -b memt-native-memory https://github.com/Brishen/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON   # or without -DGGML_CUDA=ON for CPU only
cmake --build build -j --target llama-server

To convert the original checkpoint yourself: python convert_hf_to_gguf.py <MemT-1.0-Nano dir> --outtype bf16, then llama-quantize --tensor-type memt_=q8_0 ... to keep the memory weights at Q8_0.

Native memory in llama-server

The server keeps memory states in RAM by id. Extra request fields on /v1/chat/completions and /completion:

• memory_id: memory state to read
• memory_commit: write the turns of this request into the memory when the reply ends normally
• memory_documents: list of documents to write into the memory before generating

Endpoints: GET /memory, POST /memory/{id}?action=save|restore|erase (save/restore need --slot-save-path).

./build/bin/llama-server -m MemT-1.0-Nano-Q4_K_M.gguf -ngl 99 --jinja --slot-save-path ./memory

curl localhost:8080/v1/chat/completions -d '{
  "messages": [{"role": "user", "content": "I have a three-month-old kitten named memmy who loves eating small shrimp."}],
  "memory_id": "alice", "memory_commit": true, "chat_template_kwargs": {"enable_thinking": false}}'

curl localhost:8080/v1/chat/completions -d '{
  "messages": [{"role": "user", "content": "What are my kitten'"'"'s name and favorite food?"}],
  "memory_id": "alice", "chat_template_kwargs": {"enable_thinking": false}}'

Without memory_id the model behaves as the plain backbone.

License

Apache 2.0, same as the original model. See the original model card for details, evaluation and citation.