Skip to content

Memory & Context Profiles

Full-duplex conversational voice agents often interact with users across multiple turns or multiple days. Remembering user facts (e.g. name, location, preferred language, membership tier) makes the assistant feel intelligent and personal.

edgevoiceai separates memory into: 1. Dialogue History: Raw chronological turns (ChatMessage models). 2. Entity & Profile State: Key-value facts extracted during conversations. 3. Dynamic Prompt Injection: Automatically formatting profile context into the system prompt.

1. Memory Store Implementations

edgevoiceai provides two standard implementations:

  • InMemoryMemoryStore: Ephemeral, zero-dependency in-memory dictionary. Ideal for testing and single-session interactions.
  • SQLiteMemoryStore: Lightweight file-backed SQL database. Persists conversations and user profiles across application restarts with async queries.

2. Using InMemoryMemoryStore

import asyncio
from edgevoiceai import InMemoryMemoryStore

async def example():
    memory = InMemoryMemoryStore()
    session_id = "user_42"

    # Pre-seed known profile details
    await memory.update_profile(session_id, {
        "user_name": "Rohan Sharma",
        "city": "Mumbai",
        "tier": "Premium Gold",
    })

    # Add dialogue turns
    await memory.add_message(session_id, role="user", content="Hello!")
    await memory.add_message(session_id, role="assistant", content="Hello Rohan! How can I help you today?")

    # Inspect the dynamically generated context prompt
    context_prompt = await memory.build_context_prompt(session_id)
    print(context_prompt)

asyncio.run(example())

Output: 

[KNOWN USER CONTEXT & PROFILE]
• user_name: Rohan Sharma
• city: Mumbai
• tier: Premium Gold

3. Persistent Memory with SQLiteMemoryStore

For production edge devices or servers, use SQLiteMemoryStore:

from edgevoiceai import SQLiteMemoryStore

memory = SQLiteMemoryStore(db_path="conversations.db")

The database automatically initializes the required conversations, messages, and user_profiles tables.

4. Automatic Context Injection in VoiceAgent

When you pass a memory store to VoiceAgent:

agent = VoiceAgent(
    config=config,
    tools=tools,
    memory=memory,
)

At the start of every session or turn, agent.build_effective_system_instruction(session_id) automatically calls memory.build_context_prompt(session_id) and injects the facts directly into Gemini's system instructions.