atlas/home-voice-agent/memory

Long-Term Memory System

Stores persistent facts about the user, their preferences, routines, and important information.

Features

• Persistent Storage: SQLite database for long-term storage
• Categories: Personal, family, preferences, routines, facts
• Confidence Scoring: Track certainty of each fact
• Source Tracking: Know where facts came from (explicit, inferred, confirmed)
• Fast Retrieval: Indexed lookups and search
• Prompt Integration: Format memory for LLM prompts

Usage

from memory.manager import get_memory_manager
from memory.schema import MemoryCategory, MemorySource

manager = get_memory_manager()

# Store explicit fact
manager.store_fact(
    category=MemoryCategory.PREFERENCES,
    key="favorite_color",
    value="blue",
    confidence=1.0,
    source=MemorySource.EXPLICIT
)

# Store inferred fact
manager.store_fact(
    category=MemoryCategory.ROUTINES,
    key="morning_routine",
    value="coffee at 7am",
    confidence=0.8,
    source=MemorySource.INFERRED,
    context="Mentioned in conversation on 2024-01-06"
)

# Get fact
fact = manager.get_fact(MemoryCategory.PREFERENCES, "favorite_color")
if fact:
    print(f"Favorite color: {fact.value}")

# Search facts
facts = manager.search_facts("coffee", category=MemoryCategory.ROUTINES)

# Format for LLM prompt
memory_text = manager.format_for_prompt()
# Use in system prompt: "## User Memory\n{memory_text}"

Categories

• PERSONAL: Personal facts (name, age, location)
• FAMILY: Family member information
• PREFERENCES: User preferences (favorite foods, colors)
• ROUTINES: Daily/weekly routines
• FACTS: General facts about the user

Memory Write Policy

Explicit Facts (confidence: 1.0)

• User explicitly states: "My favorite color is blue"
• Source: MemorySource.EXPLICIT

Inferred Facts (confidence: 0.7-0.9)

• Inferred from conversation: "I always have coffee at 7am"
• Source: MemorySource.INFERRED

Confirmed Facts (confidence: 0.9-1.0)

• User confirms inferred fact
• Source: MemorySource.CONFIRMED

Integration with LLM

Memory is formatted and injected into system prompts:

memory_text = manager.format_for_prompt(limit=20)
system_prompt = f"""
You are a helpful assistant.

## User Memory

{memory_text}

Use this information to provide personalized responses.
"""

Storage

• Database: data/memory.db (SQLite)
• Schema: See memory/schema.py
• Indexes: Category, key, last_accessed for fast queries

Future Enhancements

• Semantic search using embeddings
• Memory summarization
• Confidence decay over time
• Conflict resolution for conflicting facts
• Memory validation and cleanup