---
license: other
license_name: fair-noncommercial-research-license
license_link: https://huggingface.co/facebook/MobileLLM-R1-360M-base/blob/main/LICENSE
base_model: facebook/MobileLLM-R1-360M-base
language:
- en
tags:
- on-device
- executorch
- memory
- wearables
- information-extraction
---

# Remembery System 1 (v4): on-device memory for AI glasses

Two fine-tuned **MobileLLM-R1-360M** models with non-autoregressive heads that turn a live conversation transcript into
memories on the phone. Each window (~2,000 characters of transcript) gets one forward pass per model; the pooled graphs
(entities, frame triggers, roles, attributes) pick candidate lines, and a small logistic validator keeps the ones
worth remembering as structured records.

Code, app and instructions: [github.com/IAMIbrahimmemon/remembery](https://github.com/IAMIbrahimmemon/remembery).

## Scores

Held-out real conversations (Switchboard, Taskmaster, meetings); fact checklists and judging by Muse Spark.

| System | Recall | Precision | F1 |
|---|---|---|---|
| This system, sealed set (85 conversations, scored once) | 0.799 | 0.783 | **0.791** |
| This system, dev set (89 conversations) | 0.774 | 0.788 | 0.781 |
| Large LLM labeling the same dev conversations| 0.814 | 0.892 | 0.851 |

Wrong memories 0.2%. On an iPhone 17: 4.5–5.4 s per full window for both models, ~4.5 GB footprint.

## Files

| Path | What |
|---|---|
| `checkpoints/v4e/best.pt` | model A (PyTorch state dict, fp32) + `args.json`, `labels3.json` |
| `checkpoints/v4f/last.pt` | model B, same layout |
| `validator_g2.json` | line validator (logistic weights, cutoff 0.60, feature spec) |
| `vocab3.json` | pruned vocabulary used for the phone export |
| `ios/memnar-ios-model-v4.zip` | `ios/Model/`: model A as an ExecuTorch `.pte` (XNNPACK, fp32), tokenizer, metadata, parity fixtures |
| `ios/memnar-ios-model-v4-b.zip` | `ios/Model_b/`: model B, same layout |
| `ios/memnar-ios-phonon.zip` | `ios/Phonon/`: Phonon-2.5 English speech to text converted to Core ML (weights unchanged) |

Both models are needed: the score above is for the two pooled, run with cutoffs entity 0.5, trigger 0.5, role 0.35 and
the validator at 0.60.

sha256:

```
a020c56919872a17eae942ef934031a1c2125525770b752f3e58af0d31e0900b  checkpoints/v4e/best.pt
b484ad260cd4b87a4524a6a26be1e7491db06a76b0ca49d6deccdafab3786415  checkpoints/v4f/last.pt
9601602dc6f864d59ea2f2d8e7b4442947cd6d06cb3c1ecae8e7238924778196  ios/memnar-ios-model-v4.zip
7b18bbabfab1e251388534150b570b0fbe8870bb66d5d953aeb8d9614cb725f0  ios/memnar-ios-model-v4-b.zip
4dfc5b3b24ca12e7351bc5f6a3540b26d3d6439633954b834c91f751bb1f87aa  ios/memnar-ios-phonon.zip
```

## Use

```
git clone https://github.com/IAMIbrahimmemon/remembery && cd remembery && uv sync
hf download IAMIbrahim/memnar-system1 --include "checkpoints/*" --local-dir hf
PYTHONPATH=. uv run python memories.py --ckpts hf/checkpoints --windows data/probe/probe_v1.jsonl --out memories.jsonl
```

## Licenses

- System 1 checkpoints and `.pte` files: derivatives of MobileLLM-R1-360M-base, under the FAIR Noncommercial Research
  License (noncommercial research use only).
- `ios/memnar-ios-phonon.zip`: Phonon-2.5 English by FermionResearch (a compressed nvidia/parakeet-tdt-0.6b-v3),
  CC-BY-4.0, converted to Core ML without changing the weights.

Built for the Meta Global AI Developer Hackathon 2026.