---
license: mit
library_name: convmemory
tags:
- retrieval
- memory
- reranking
- agents
- convmemory
- ccge-la
pipeline_tag: feature-extraction
---

# ConvMemory CCGE-LA LoCoMo MPNet Seed-23 Alpha

This repository contains an alpha CCGE-LA conflict editor checkpoint for the public ConvMemory API.

CCGE-LA stands for **Low-Amplitude Counterfactual Conflict Graph Editor**. It is a lightweight post-ConvMemory editor for stale/current memory conflicts:

```text
vector search -> ConvMemory -> CCGE-LA conflict-aware score edit -> memory context
```

## Files

- `ccge_la.pt`: CCGE-LA editor checkpoint.
- `manifest.json`: training configuration and seed-23 test metrics.
- `LICENSE`: MIT license.

## Usage

Install ConvMemory from GitHub, or from PyPI after the next package release:

```bash
pip install git+https://github.com/pth2002/ConvMemory.git
```

Load both checkpoints directly from Hugging Face Hub:

```python
from convmemory import ConvMemory

model = ConvMemory.from_pretrained("Purdy0228/ConvMemory-LoCoMo-MPNet")
model.load_ccge_editor("Purdy0228/ConvMemory-CCGE-LA")

results = model.retrieve(
    query=query,
    memories=memories,
    editor="ccge_la",
    top_k=10,
)
```

For systems with precomputed embeddings, skip encoder loading on the base model:

```python
model = ConvMemory.from_pretrained("Purdy0228/ConvMemory-LoCoMo-MPNet", embedding_model=False)
model.load_ccge_editor("Purdy0228/ConvMemory-CCGE-LA")
ranked = model.rerank_embeddings(
    query_embedding=query_embedding,
    memory_embeddings=memory_embeddings,
    memory_ids=memory_ids,
    memory_texts=memory_texts,
    query=query,
    editor="ccge_la",
)
```

## Metrics

These are seed-23 test metrics from the release manifest. This is an alpha checkpoint, not a final benchmark release.

| subset | CCGE-LA alpha MRR | CCGE-LA R@10 | gate |
|---|---:|---:|---:|
| FULL | 0.5638 | 0.7725 | 0.0995 |
| T_SUP_auto | 0.5508 | 0.7138 | 0.0995 |
| CONV_TOP1_WRONG_GOLD_IN_POOL | 0.2994 | 0.6822 | 0.0995 |
| RESCUABLE_STALE_TOP1 | 0.3093 | 0.6877 | 0.0995 |

## Training Notes

- Base checkpoint: [`Purdy0228/ConvMemory-LoCoMo-MPNet`](https://huggingface.co/Purdy0228/ConvMemory-LoCoMo-MPNet).
- Training split seed: `23`.
- Candidate top-n: `192`.
- Objective: retrieval cross-entropy plus a low-amplitude gate budget penalty.
- No current/stale labels, no gold-defined feature, and no distillation objective are used by the editor.

## Limitations

- This is a public alpha checkpoint trained on a single LoCoMo-style seed-23 split.
- It is intended for API trials and early integration, not as a final benchmark claim.
- It should be used with the matching MPNet-family ConvMemory checkpoint.
- No inference widget is provided; use the `convmemory` Python library.

## Citation

A formal citation will be added when a technical report is available.

## Links

- GitHub: https://github.com/pth2002/ConvMemory
- Base ConvMemory checkpoint: https://huggingface.co/Purdy0228/ConvMemory-LoCoMo-MPNet
- CCGE-LA docs: https://github.com/pth2002/ConvMemory/blob/main/docs/CCGE_LA.md