| --- |
| license: apache-2.0 |
| base_model: |
| - Qwen/Qwen3-0.6B |
| - Qwen/Qwen3-1.7B |
| language: en |
| tags: [onnx, onnxruntime-genai, memory, notes, sourcedrecall] |
| --- |
| |
| # sourcedrecall notes models (English, int4 ONNX) |
|
|
| Two small models for one job: given one conversation, write the lasting |
| facts the named person states about themselves, one short sentence per line, |
| or NONE. sourcedrecall runs them inside its own process on onnxruntime-genai, |
| on the CPU, with no server. |
|
|
| | file | base | download | a chat session (4 CPU threads) | peak memory | |
| |---|---|---|---|---| |
| | `sourcedrecall-notes-en-0.6b.tar.gz` (small) | Qwen3-0.6B | 0.51 GB | 2-3 s | 1.9 GB | |
| | `sourcedrecall-notes-en-1.7b.tar.gz` (standard) | Qwen3-1.7B | 1.3 GB | about 4.5 s | 3.3 GB | |
|
|
| A long coding session (eight parts of 12,000 characters) takes about 100 s |
| and up to 3.6 GB with the small model; sourcedrecall writes notes in the |
| background after the session ends. |
|
|
| They are optional. sourcedrecall stores and finds the user's own words |
| without them; with one, short notes are added, each kept only if its words |
| are in what the user wrote, and shown labelled as written by a model. |
|
|
| pip install 'sourcedrecall[notes]' |
| sourcedrecall-setup --notes # small |
| sourcedrecall-setup --notes standard # standard |
| |
| ## How they were made |
|
|
| - Full fine-tune of the base model with mlx-lm, then int4 ONNX with the |
| onnxruntime-genai builder (k_quant_mixed). |
| - Teacher: Qwen3-14B wrote notes for 3,000 LongMemEval-S conversations (MIT) |
| and 1,500 SODA dialogues (CC BY 4.0), each SODA dialogue from both |
| speakers' sides; a note was kept only if most of its words are in the |
| person's own lines. |
| - Every training conversation got a random owner name and date format, and |
| its pronouns were made explicit (I/my -> the speaker, you/your -> the |
| listener), as sourcedrecall does before the model reads a conversation. |
| - Scripts: `tools/distil` in the sourcedrecall repository. |
|
|
| ## Measured (local Qwen3-14B reader and judge for every system) |
|
|
| LoCoMo, conversations 2-9 (1,307 questions; used to choose between models, so not held out): |
|
|
| | | overall | multi-hop | temporal | single-hop | |
| |---|---|---|---|---| |
| | sourcedrecall, no notes | 64.7% | 52.3% | 53.5% | 75.5% | |
| | with the small model's notes | 68.2% | 55.2% | 58.9% | 79.0% | |
| | with the standard model's notes | 70.3% | 59.0% | 57.4% | 81.6% | |
| | with Qwen3-14B's notes (a server) | 73.1% | 62.3% | 57.4% | 85.4% | |
| | Mem0 2.2.1 | 64.6% | 61.9% | 37.2% | 77.9% | |
|
|
| Notes on the LoCoMo development conversations, judged against their |
| conversation by Qwen3-14B: said by the person 80% (small), 87% (standard), |
| 93% (Qwen3-14B); gave the person the other speaker's facts 13%, 9%, 3%. |
| Held-out false-memory sets: v3 0 of 30 for both; v4, an old state given as |
| current, 3 of 44 (small) and 4 of 44 (standard), against 3 with Qwen3-14B's |
| notes and 2 without notes. |
|
|
| These answer scores come from local models and are not comparable with |
| published LoCoMo results, which use GPT-4-class models. |
|
|
| ## Limits |
|
|
| English only. They sometimes give the user something the other person said |
| ("has kids" from the other person's children), or keep a note whose words |
| the user wrote but whose meaning differs (a negation, a number). The small |
| model often writes nothing for short sessions; standard does better there. |
| Neither closes sourcedrecall's multi-hop gap to Mem0. The held-out false |
| memory sets were run many times during development; v4 is no longer clean. |
| Known issues and fixes in progress: docs/REVIEW_2026-10-09.md in the |
| sourcedrecall repository. |
| |
| ## Attribution |
| |
| SODA (Kim et al., 2023), CC BY 4.0. LongMemEval (Wu et al., 2024), MIT. |
| Qwen3 (Qwen team), Apache-2.0. |
| |
