Memory Settings

Configure the memory system for personalized AI interactions.

Enable Memory

Master toggle for the entire memory system.

When disabled:

• No memories are retrieved
• No new memories are created
• Existing memories are preserved

Retrieval Settings

Auto-Retrieval

Automatically retrieve relevant memories before AI responses.

How it works:

1. Your message is analyzed
2. Similar memories are found using semantic search
3. Relevant context is provided to the AI

Query Rewriting

When enabled, the AI optimizes your search queries before searching memories.

This improves retrieval accuracy by reformulating queries to better match how memories are stored.

Max Retrieved Memories

Maximum number of memories to retrieve per message (1-20).

More memories provide more context but may slow responses.

Similarity Threshold

Minimum relevance score for memory retrieval (0-100%).

• 50%: More memories, less relevant (loose match)
• 70%: Balanced (recommended)
• 90%: Fewer memories, highly relevant (strict match)

Summarization Settings

Auto-Summarization

Automatically extract memories from conversations.

After each conversation:

1. Important information is identified
2. Memories are created or updated
3. Stored with semantic embeddings for later retrieval

Memory Tool Model

The Memory Tool Model handles memory-specific AI operations. By default, it uses the global Tool Model, but you can configure a separate model for memory operations.

What Memory Tool Model Does

• Memory Extraction: Identifying important information to remember
• Relevance Analysis: Determining which memories are relevant
• Query Understanding: Interpreting your messages to find matching memories

Configuration

You have two options:

1. Use Default Tool Model: Memory operations use the global Tool Model configured in General settings. This is recommended for most users.

2. Custom Memory Model: Configure a specific model just for memory operations. Useful if you want:

   • A different cost/quality tradeoff for memory vs. tools
   • To use a model optimized for text understanding

DANGER

Never use reasoning models (o1, o3, extended thinking) for memory operations - they are far too slow.

Embedding Model

Model used for creating memory embeddings (vector representations for semantic search).

Supported Models

• text-embedding-3-small (OpenAI) - Default, good balance
• text-embedding-3-large (OpenAI) - Higher quality
• Google Embedding API
• Custom OpenAI-compatible embedding APIs

Changing Models

When you change the embedding model:

1. A warning will appear if you have existing memories
2. Click Rebuild Embeddings to re-process all memories
3. Wait for the rebuild to complete

WARNING

Rebuilding embeddings can take time for large memory databases. You can cancel the rebuild if needed.

Memory Statistics

View memory system statistics:

• Total Memories: Count of all stored memories
• Auto Generated: Memories created automatically from conversations
• Manually Added: Memories you've added directly

Add Memory

Manually add memories that the AI should remember:

1. Enter the memory content in the text area
2. Click Add Memory
3. The memory is stored with semantic embeddings

Use this for important information you want the AI to always have access to.

Search Memories

Test memory retrieval by searching:

1. Enter a search query
2. Click the search button
3. View matching memories with relevance scores

If Query Rewriting is enabled, you'll see how your query was optimized.

Memory List

Browse, edit, and manage all your memories:

• Refresh: Reload the memory list
• Edit: Click the edit icon to modify a memory
• Delete: Click the delete icon to remove a memory
• Clear All: Delete all memories (use with caution)

Each memory shows:

• Content
• Source (auto or manual)
• Durability (permanent or temporary)
• Creation date
• Access count (how often it's been retrieved)
• Associated thread (if from a conversation)

DANGER

Clearing all memories cannot be undone. Consider if you really need to delete everything.