---
pretty_name: PraxisMem-Training
license: other
license_name: praxismem-training-data-terms
license_link: LICENSE
language:
- en
tags:
- memory
- reinforcement-learning
---

# PraxisMem-Training

Training data for PraxisMem's memory update and retrieval agents, including
VERL-compatible samples and canonical memory states.

## Data

Download [verl.tar.gz](verl.tar.gz) and extract it into your data directory.

| Task | Samples |
|---|---:|
| Memory update | 481 |
| Memory retrieval | 1,440 |

Each sample has a unique `sample_id`. Multiple retrieval samples may share
the same trajectory, session, and checkpoint, with their own query and target.

```text
verl/
  README.md
  LICENSE
  opsd/
    update.parquet
    retrieval.parquet
    samples.jsonl
  canonical_states/
    manifest.json
    update_sessions-0000{0,1,2}.jsonl
    update_chunks-00000.jsonl
```

Canonical state shards contain 482 sessions and 3,375 ordered chunk deltas.
Use `verl/` as the dataset directory and `verl/canonical_states/` as
`canonical_states_root` in the companion training configuration.

For setup, sample loading, and training instructions, see the
[PraxisMem training documentation](https://github.com/QMMMS/PraxisMemBench/tree/main/docs/training).
The companion [PraxisMemBench dataset](https://huggingface.co/datasets/QMMMS/PraxisMemBench)
provides benchmark trajectories and conversation history.

## License And Sources

See [LICENSE](LICENSE) for data sources and component-specific terms.