Bot Memory Management
Memoh's structured long-term memory system allows Bots to remember information across multiple conversations, providing contextually relevant and personalized interactions.
Concept: Semantic Search
Memories are stored in a Qdrant vector database. When a user asks a question, Memoh performs a semantic search to find the most relevant "memories" and includes them in the bot's system prompt.
Operations
Manage your bot's memories from the Memory tab in the Bot Detail page.
1. Creating Memories
- New Memory: Manually enter a memory's content in the provided textarea.
- From Conversation: Select specific messages from the bot's conversation history to "extract" into memory.
2. Searching and Managing
- Search: Filter memories by ID or text content.
- Edit: Modify existing memory entries directly in the list.
- Delete: Remove memories that are no longer needed.
Memory Compression (Compact)
Over time, memories can accumulate and become redundant. The Compact feature helps optimize the memory pool.
- Ratio: Set the compression ratio (e.g.,
0.8,0.5,0.3) to determine how much information is retained. - Decay Days: Optionally specify a time window to only compact memories older than a certain number of days.
Visualization: Vector Manifold
The Memory tab includes advanced visual tools to help you understand how the memory system is performing:
Top-K Bucket Chart
Shows the distribution of relevant memories retrieved for the most recent queries.
CDF Curve (Cumulative Distribution Function)
Visualizes the scoring threshold of retrieved memories, helping you fine-tune how much "relevant" information the bot should consider.
Bot Interaction
- The bot will automatically search for and retrieve memories during every interaction.
- The Memory Model configured in the Settings tab is used for extracting and summarizing these memories.
- Memories provide the "long-term knowledge" that makes each bot unique to its owner.