A semantic memory tool for OpenWebUI that enables AI assistants to store, retrieve, and manage user-specific memories using vector embeddings and natural language processing.
- Semantic storage and retrieval of user facts
- Multilingual support (stores in English)
- Automatic deduplication and updating
- LRU caching and async operations for speed
- User privacy: local, isolated processing
Uses a configurable sentence transformer (default: Alibaba-NLP/gte-multilingual-base) to create embeddings for memories and queries. The model can be changed via the SENTENCE_TRANSFORMER_MODEL valve parameter.
- Extracts important facts from conversations
- Converts text to vector embeddings
- Prevents duplicates and updates similar memories
- Retrieves relevant memories based on meaning
- Upload
memory.pyin OpenWebUI (Workspace → Tools) - Adjust valve parameters if needed (defaults work well)
- Enable the tool
No restart required – tools are loaded dynamically
| Parameter | Default | Description |
|---|---|---|
USE_MEMORY |
true |
Enable/disable memory features |
SIMILARITY_BASE_THRESHOLD |
0.80 |
Base similarity score for memory matching |
RETRIEVE_THRESHOLD_MULTIPLIER |
0.7 |
Lower = more inclusive retrieval |
MAX_RELEVANT_CONTEXT_RESULTS |
20 |
Maximum memories returned per query |
SENTENCE_TRANSFORMER_MODEL |
Alibaba-NLP/gte-multilingual-base |
HuggingFace model name for SentenceTransformer |
Tip: Lower RETRIEVE_THRESHOLD_MULTIPLIER for more inclusive retrieval. You can set SENTENCE_TRANSFORMER_MODEL to any compatible HuggingFace model name.
- Stores: Important facts (profession, family, preferences)
- Retrieves: Relevant memories for user queries
- Updates: Existing memories with new info
Stored: Personal facts, relationships, profession, preferences, important dates
Not stored: Temporary info, opinions, daily tasks, common knowledge
- Check
USE_MEMORYis enabled - Lower
RETRIEVE_THRESHOLD_MULTIPLIERfor more results - Use descriptive queries
Built for the OpenWebUI community 🚀