kurcontko/mnemotree-leaf-v1

{"fact": "Python uses GIL for thread safety", "subject": "Python", "confidence": 0.92}

{"event": "Alice went hiking in Yosemite", "who": "Alice", "confidence": 0.88}

{"procedure": "Deploy with docker compose up -d", "subject": "deployment", "confidence": 0.85}

from transformers import AutoTokenizer, AutoModelForCausalLM
import json

model = AutoModelForCausalLM.from_pretrained(
    "kurcontko/mnemotree-leaf-v1", torch_dtype="bfloat16", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("kurcontko/mnemotree-leaf-v1")

system = (
    "You are a memory extraction assistant. Given a conversation turn with a "
    "type prefix (<|semantic|>, <|episodic|>, or <|procedural|>), extract "
    "structured information as a JSON object. Output ONLY valid JSON, no explanation."
)

user = "<|semantic|> Python uses the GIL for thread safety in CPython."

messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": user},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(inputs, max_new_tokens=256, temperature=0.1, top_p=0.95)

result = tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True)
parsed = json.loads(result)
# {"fact": "Python uses GIL for thread safety in CPython", "subject": "Python", "confidence": 0.92}

from mnemotree import MemoryCoreBuilder

memory = (
    MemoryCoreBuilder(store)
    .with_local_models(device="cuda")  # auto-downloads root + leaf
    .build()
)

item = await memory.remember("Python uses the GIL for thread safety")
print(item.metadata["extraction"])
# {"fact": "...", "subject": "Python", "confidence": 0.92}

root (5ms) → classify type → leaf (1s) → extract structured JSON

@misc{mnemotree2025,
  title={mnemotree: Local-first memory for LLM agents},
  author={kurcontko},
  year={2025},
  url={https://github.com/kurcontko/mnemotree}
}