---
license: other
license_name: tongyi-qianwen-research
license_link: https://huggingface.co/Qwen/Qwen1.5-1.8B/blob/main/LICENSE
base_model: Qwen/Qwen1.5-1.8B
tags:
  - gguf
  - qwen
  - qwen1.5
  - lora
  - memory
  - agent-memory
  - psm-memory
library_name: transformers
pipeline_tag: text-generation
---

# PSM Memory Qwen 1.5B

PSM Memory is a small model trained for memory-management operations in agent systems. It is not intended to answer user questions directly. It is intended to decide what should be remembered, retrieve and rank relevant memories, detect conflicts, and support memory lifecycle operations.

This repository contains the runtime artifacts for the PSM Memory model derived from Qwen1.5-1.8B.

## Files

### GGUF runtime artifacts

- `gguf/psm-memory-qwen-1.5b-q4_k_m.gguf`
  - Recommended default runtime artifact for `@psm-memory/cli` and `@psm-memory/pi-plugin`.
  - Quantized GGUF for local inference.

- `gguf/psm-memory-qwen-1.5b-f16.gguf`
  - Full precision GGUF export.
  - Larger and slower, useful for quality comparison and future quantization.

### Hugging Face / Transformers artifacts

- `merged-fp16/`
  - Merged FP16 model weights and tokenizer/config files.
  - Useful for inspection, conversion, or further export.

- `lora/`
  - LoRA adapter weights and tokenizer/config files.
  - Useful for reproducibility and future fine-tuning work.

### Integrity

- `checksums.sha256`
  - SHA-256 checksums for uploaded artifacts.

## Intended Use

Use this model as a memory-management component for an LLM agent:

1. User prompt arrives.
2. PSM Memory retrieves/ranks relevant memories and produces context for the main LLM.
3. Main LLM responds or takes an action.
4. PSM Memory evaluates the response/action and stores durable memory when appropriate.

The npm packages are:

- `@psm-memory/sdk`
- `@psm-memory/cli`
- `@psm-memory/pi-plugin`

The CLI/plugin should use the Q4_K_M GGUF artifact by default.

## Not Intended For

- General chat completion.
- Factual QA as a standalone assistant.
- Safety-critical autonomous decision-making without external validation.
- Storing private user data without consent and appropriate access controls.

## License

This model is a fine-tuned/converted PSM Memory model derived from Qwen1.5-1.8B.

The model weights are distributed under the Tongyi Qianwen Research License inherited from the base model:

https://huggingface.co/Qwen/Qwen1.5-1.8B/blob/main/LICENSE

The PSM Memory npm package code is licensed separately in its source repository.

## Training Data

The upload bundle does not include training data by default. If synthetic training data is published, it should be uploaded only after checking that it contains no private, personal, or proprietary content.

Recommended approach:

- Add a separate `training-data/` folder for synthetic examples.
- Include a short data card explaining how the data was generated.
- Include only sanitized synthetic data or a representative sample if the full dataset is large.

## Reproducibility Notes

A Colab notebook is not required for using this model. It is useful if you want others to reproduce conversion, quantization, or adapter loading.

Recommended optional notebooks:

- `notebooks/convert-to-gguf.ipynb`
- `notebooks/quantize-gguf.ipynb`
- `notebooks/run-psm-memory.ipynb`

These can be added later without blocking CLI setup support.