Memory System

Alma's memory system allows the AI to remember important information across conversations, creating a more personalized and context-aware experience.

How It Works

1. Extraction: During conversations, Alma identifies important information
2. Storage: Memories are stored with semantic embeddings for efficient retrieval
3. Retrieval: Before each response, relevant memories are automatically retrieved
4. Usage: The AI uses these memories to provide more contextual responses

Types of Memories

Automatic Memories

Extracted automatically from conversations:

• User preferences
• Important facts
• Project details
• Technical knowledge

Manual Memories

Created by you or explicitly requested:

• Click "Remember this" on any message
• Ask the AI to "Remember that..."
• Create memories in Settings → Memory

Memory Settings

Enable Memory

Master toggle for the entire memory system.

Auto-Summarization

Automatically extracts memories from conversations after they end.

Auto-Retrieval

Automatically retrieves relevant memories before generating responses.

Retrieval Settings

Setting	Description	Default
Max Retrieved	Maximum memories to retrieve	10
Similarity Threshold	Minimum relevance score (0-1)	0.7
Query Rewriting	Optimize queries for better retrieval	On

Memory Management

Viewing Memories

1. Go to Settings → Memory
2. Click View All Memories
3. Browse, search, and filter your memories

Editing Memories

Click any memory to:

• Edit the content
• Change importance
• Add or remove tags
• Toggle durability

Deleting Memories

• Delete individual memories from the memory list
• Use Clear All to remove all memories (caution!)

Memory Properties

Importance (0-1)

Higher importance memories are retrieved more readily.

Durability

• Permanent: Never automatically deleted
• Temporary: May be pruned during maintenance

Tags

Organize memories with custom tags for easier filtering.

Source

Tracks where the memory came from:

• Chat (extracted from conversation)
• Manual (created by user)
• Auto (created by system)

Embedding Models

Memories use vector embeddings for semantic search. Supported models:

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

Changing Embedding Model

1. Go to Settings → Memory
2. Select a different embedding model
3. Click Rebuild Embeddings

WARNING

Rebuilding embeddings processes all memories and may take time for large databases.

Incognito Mode

When you want a conversation without memory:

1. Enable Incognito Mode in the chat toolbar
2. No memories will be retrieved or created
3. The conversation is excluded from memory processing

Best Practices

What to Remember

• Personal preferences (coding style, writing tone)
• Project-specific information
• Important decisions and rationale
• Frequently referenced facts

What NOT to Remember

• Temporary or one-time information
• Sensitive data (passwords, keys)
• Information that changes frequently

Memory Hygiene

• Periodically review and clean up old memories
• Merge duplicate or similar memories
• Update outdated information

Troubleshooting

"Memories Not Retrieved"

• Check if memory is enabled
• Verify similarity threshold isn't too high
• Ensure auto-retrieval is on

"Wrong Memories Retrieved"

• Lower the similarity threshold
• Add more specific tags
• Improve memory content clarity

"Too Many Memories"

• Use tags to organize
• Delete irrelevant memories
• Increase similarity threshold