mnemotree-root-v1

Fast memory type classifier for mnemotree — a local-first memory system for LLM agents.

Classifies conversation turns into memory types: episodic, semantic, or procedural. Runs in ~5ms on GPU, ~15ms on CPU.

Model Details

Base model	ModernBERT-base (149M params)
Task	Multi-label sequence classification (3 labels)
Training	Full fine-tune, Focal Loss (gamma=2.0)
Precision	bfloat16
Max seq length	256 tokens
Size	289 MB

Performance

Metric	Score
Macro F1	0.674
Micro F1	0.703
Optimal threshold	0.55

Per-class metrics (threshold=0.55)

Class	Precision	Recall	F1
Episodic	0.627	0.668	0.647
Semantic	0.810	0.690	0.745
Procedural	0.552	0.737	0.631

Label Mapping

0 → episodic    (events, experiences, conversations)
1 → semantic    (facts, knowledge, definitions)
2 → procedural  (how-to, workflows, instructions)

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained("kurcontko/mnemotree-root-v1")
tokenizer = AutoTokenizer.from_pretrained("kurcontko/mnemotree-root-v1")

text = "Python uses the GIL for thread safety in CPython."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits)[0]

labels = ["episodic", "semantic", "procedural"]
for label, prob in zip(labels, probs):
    print(f"{label}: {prob:.3f}")
# episodic: 0.12, semantic: 0.94, procedural: 0.08

With mnemotree

from mnemotree import MemoryCoreBuilder

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

await memory.remember("Python uses the GIL for thread safety")
# → classified as semantic, importance=0.94

Training

• Dataset: 73,781 samples (49,706 train / 10,024 val / 14,051 test)
• Sources: LoCoMo, MSC, code synthesis, nameswap/paraphrase augmentations
• Epochs: 5
• Batch size: 64
• Learning rate: 2e-5 (cosine schedule, 10% warmup)
• Loss: Binary Cross-Entropy with Focal Loss (gamma=2.0)
• No conversation leakage between splits

Companion Model

Pair with mnemotree-leaf-v1 (SmolLM2-1.7B extractor) for structured memory extraction.

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

Citation

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