evrmem Skill\r

\r

Name\r

\r evrmem\r \r

Description\r

\r Local Chinese Vector Memory System. Provides semantic memory search and storage for AI agents using local Chinese embedding models (text2vec) and ChromaDB. Supports RAG-based context augmentation.\r \r

When to Use\r

\r Use this skill when the user asks to:\r

• "Search memories" or "Find related memories"\r
• "Save this to memory"\r
• "Remember this information"\r
• "Search my knowledge base"\r
• "Find past notes about X"\r
• "Add this to my memory"\r
• "What do I know about X"\r
• "RAG retrieval" or "context augmentation"\r
• Query or recall previous learnings\r \r

Prerequisites\r

\r Install evrmem and initialize:\r \r

pip install evrmem\r
evrmem init\r
```\r
\r
For China users (mirror):\r
\r
```bash\r
set HF_ENDPOINT=https://hf-mirror.com   # Windows\r
# or\r
export HF_ENDPOINT=https://hf-mirror.com   # Linux/Mac\r
evrmem init\r
```\r
\r
## Core Workflow\r
\r
### 1. Semantic Search (Most Common)\r
\r
```python\r
from qmd.core.vector_db import vector_db\r
\r
results = vector_db.search("React form warning", top_k=5)\r
for r in results:\r
    print(f"[{r['distance']:.3f}] {r['content'][:80]}")\r
```\r
\r
Or via CLI:\r
\r
```bash\r
evrmem search "React form warning"\r
evrmem search "deployment issue" --project myproject\r
```\r
\r
### 2. Add Memory\r
\r
```python\r
memory_id = vector_db.add_memory(\r
    "React StrictMode causes Form.useForm warning",\r
    metadata={"project": "mes-demo", "tags": "react,antd"}\r
)\r
```\r
\r
Or via CLI:\r
\r
```bash\r
evrmem add "Important finding about X" --project myproject --tags react,bug\r
```\r
\r
### 3. Structured Query\r
\r
```bash\r
# Query by project\r
evrmem query --project mes-demo\r
\r
# Query by tag\r
evrmem query --tag react\r
\r
# List all projects\r
evrmem query --list-projects\r
\r
# List all tags\r
evrmem query --list-tags\r
```\r
\r
### 4. RAG Retrieval\r
\r
```python\r
result = vector_db.rag("how to fix the form warning", top_k=3)\r
print(result["context"])\r
```\r
\r
Or via CLI:\r
\r
```bash\r
evrmem rag "how to fix the form warning"\r
evrmem rag "how to fix the form warning" --prompt\r
```\r
\r
### 5. Statistics\r
\r
```bash\r
evrmem stats\r
```\r
\r
## Configuration\r
\r
Create `~/.evrmem/config.yaml`:\r
\r
```yaml\r
vector_db:\r
  persist_directory: "~/.evrmem/data/qmd_memory"\r
\r
embedding:\r
  model_name: "shibing624/text2vec-base-chinese"\r
  device: "cpu"  # or "cuda"\r
  cache_folder: "~/.evrmem/models"\r
\r
rag:\r
  top_k: 5\r
  min_similarity: 0.5\r
\r
logging:\r
  level: "WARNING"\r
```\r
\r
## Environment Variables\r
\r
| Variable | Description | Default |\r
|----------|-------------|---------|\r
| `EVREM_DATA_DIR` | Data directory | `~/.evrmem/data/qmd_memory` |\r
| `EVREM_MODEL_NAME` | HuggingFace model name | `shibing624/text2vec-base-chinese` |\r
| `EVREM_LOCAL_MODEL` | Local model path (highest priority) | - |\r
| `EVREM_DEVICE` | Device for inference | `cpu` |\r
| `EVREM_TOP_K` | Default retrieval count | `5` |\r
| `EVREM_MIN_SIM` | Minimum similarity threshold | `0.5` |\r
| `EVREM_LOG_LEVEL` | Logging level | `WARNING` |\r
| `EVREM_LOCAL_FILES_ONLY` | Disable network access | `false` |\r
| `HF_ENDPOINT` | HuggingFace mirror endpoint | - |\r
\r
## Response Format\r
\r
When reporting search results, use this format:\r
\r
```\r
## evrmem Search Results\r
\r
**Query:** "user query"\r
**Results:** N memories found\r
\r
| Score | Project | Content |\r
|-------|---------|---------|\r
| 0.723 | mes-demo | React StrictMode causes Form.useForm warning... |\r
| 0.681 | docs | Deployment script timeout issue... |\r
\r
### Top Match\r
**Project:** mes-demo | **Tags:** react,antd\r
\r
> React StrictMode causes Form.useForm warning...\r
```\r
\r
When adding memory:\r
\r
```\r
## Memory Saved\r
\r
**ID:** abc123\r
**Project:** mes-demo\r
**Tags:** react\r
**Content:** React StrictMode causes Form.useForm warning...\r
\r
Use `evrmem search "React StrictMode"` to retrieve later.\r
```\r
\r
## Installation for Agent\r
\r
If `evrmem` is not installed:\r
\r
```python\r
import subprocess\r
subprocess.run(["pip", "install", "evrmem"], check=True)\r
# Initialize on first use (downloads ~400MB model)\r
subprocess.run(["evrmem", "init"], check=True)\r
```\r
\r
For China users, set mirror before init:\r
\r
```python\r
import os\r
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"\r
subprocess.run(["evrmem", "init"], check=True)\r
```\r
\r
## Edge Cases\r
\r
- **Model download fails**: Set `HF_ENDPOINT=https://hf-mirror.com` before `evrmem init`\r
- **NumPy errors**: Run `pip install "numpy\x3C2" --force-reinstall`\r
- **Offline/air-gapped**: Download model on connected machine, copy `~/.evrmem/models` to offline machine, set `EVREM_LOCAL_FILES_ONLY=true`\r
- **Empty search results**: Try broader terms or check if memories exist with `evrmem query --list-projects`\r
- **Similarity too low**: Adjust `--top-k` or lower `EVREM_MIN_SIM` threshold\r
- **Slow search**: Use CPU by default; set `EVREM_DEVICE=cuda` if GPU available\r