Can LLMs create lasting flashcards from readers' highlights?
23
stars
6
commits
Python
primary language
Jul 1, 2026
updated
README
memory-machines
Research code for evaluating and training LLM-based memory prompt (flashcards) generators for spaced repetition.
Datasets
Both datasets are published under the laddermedia HuggingFace org.
| Dataset | Granularity | Purpose |
|---|---|---|
laddermedia/srs-prompts | one row per candidate memory prompt | Card-level: pluckability classification, reward-model preference pairs, raw input for srs-highlights |
laddermedia/srs-highlights | one row per highlight (with all candidate prompts grouped) | Highlight-level: tiering evaluation, SFT tiering, masked task tiering |
srs-highlights is built from srs-prompts via memory_machines.highlight.build_dataset.
Setup
uv sync
Citation
If you use this code or the associated datasets, please cite:
@misc{memory-machines,
title = {Memory Machines: Can LLMs create lasting flashcards from readers' highlights?},
author = {Kirkby, Ozzie and Matuschak, Andy},
year = {2026},
address = {San Francisco},
url = {https://memory-machines.com/}
}
License
Apache-2.0. See LICENSE.
Contributors
6 commits
Similar repos
Languages
Python
92.9%
Shell
6.4%
Can LLMs create lasting flashcards from readers' highlights?
23
stars
6
commits
Python
primary language
Jul 1, 2026
updated
README
memory-machines
Research code for evaluating and training LLM-based memory prompt (flashcards) generators for spaced repetition.
Datasets
Both datasets are published under the laddermedia HuggingFace org.
| Dataset | Granularity | Purpose |
|---|---|---|
laddermedia/srs-prompts | one row per candidate memory prompt | Card-level: pluckability classification, reward-model preference pairs, raw input for srs-highlights |
laddermedia/srs-highlights | one row per highlight (with all candidate prompts grouped) | Highlight-level: tiering evaluation, SFT tiering, masked task tiering |
srs-highlights is built from srs-prompts via memory_machines.highlight.build_dataset.
Setup
uv sync
Citation
If you use this code or the associated datasets, please cite:
@misc{memory-machines,
title = {Memory Machines: Can LLMs create lasting flashcards from readers' highlights?},
author = {Kirkby, Ozzie and Matuschak, Andy},
year = {2026},
address = {San Francisco},
url = {https://memory-machines.com/}
}
License
Apache-2.0. See LICENSE.
Similar repos
Contributors
6 commits
Languages
Python
92.9%
Shell
6.4%