Memory in Marvin allows agents to store and recall information across multiple conversations and sessions. Unlike threads, which maintain context within a single interaction, memory provides long-term persistence that survives across program restarts.

​
What is Memory?

A Memory in Marvin is a specialized module that:

  1. Stores information in a vector database for efficient retrieval
  2. Allows agents to remember facts, preferences, and knowledge
  3. Can be shared between different agents or kept private
  4. Persists across multiple conversations and program executions
import marvin

# Create a memory module
user_preferences = marvin.Memory(
    key="user_preferences",
    instructions="Remember the user's preferences and settings"
)

# Create an agent with this memory
assistant = marvin.Agent(
    name="Personal Assistant",
    memories=[user_preferences]
)

# The agent can now store and recall information in this memory
assistant.run("Remember that I prefer dark mode in all applications")

# Later (even in a different session), the agent can recall this information
assistant.run("What are my UI preferences?")  # Will recall dark mode preference

​
Memory vs. Thread

It’s important to understand the difference between Memory and Thread:

FeatureThreadMemory
PersistenceTemporary (session)Long-term (across sessions)
StorageSQLite databaseVector database
ScopeAll messages in a conversationSpecific information to remember
RetrievalSequential accessSemantic search
PurposeMaintain conversation flowStore specific knowledge

​
Creating and Using Memory

​
Basic Memory Creation

To create a memory module:

import marvin

memory = marvin.Memory(
    key="project_knowledge",
    instructions="Store information about our software project"
)

The parameters are:

  • key: A unique identifier for this memory (alphanumeric and underscores only)
  • instructions: Guidance for how the memory should be used

​
Attaching Memory to Agents

Memory modules are attached to agents:

# Create a memory
product_knowledge = marvin.Memory(key="product_specs")

# Attach to an agent
support_agent = marvin.Agent(
    name="Support Agent",
    instructions="You provide technical support",
    memories=[product_knowledge]
)

# Use the agent with memory
support_agent.run("Remember that our product has 8GB RAM and 256GB storage")
support_agent.run("What are our product specifications?")

​
Sharing Memory Between Agents

Multiple agents can share the same memory:

# Create a shared memory
team_knowledge = marvin.Memory(key="team_knowledge")

# Create agents that share this memory
researcher = marvin.Agent(
    name="Researcher",
    memories=[team_knowledge]
)

writer = marvin.Agent(
    name="Writer",
    memories=[team_knowledge]
)

# The researcher stores information
researcher.run("Remember that quantum computing uses qubits instead of bits")

# The writer can access this information
writer.run("Write a paragraph about quantum computing fundamentals")

​
Memory Operations

​
Adding Information

Memory is typically populated through agent interactions, but you can also add information programmatically:

import marvin
import asyncio

async def main():
    memory = marvin.Memory(key="documentation")
    
    # Add information to memory
    await memory.add("Python is a high-level programming language known for readability")
    await memory.add("FastAPI is a modern web framework for building APIs with Python")

asyncio.run(main())

​
Searching Memory

You can search memory directly:

import marvin
import asyncio

async def main():
    memory = marvin.Memory(key="documentation")
    
    # Search for relevant information
    results = await memory.search("What is Python?")
    print(results)  # Dictionary of memory_id: content pairs

asyncio.run(main())

​
Deleting Memory

To remove specific memories:

import marvin
import asyncio

async def main():
    memory = marvin.Memory(key="user_data")
    
    # Add some information
    memory_id = await memory.add("User prefers dark mode")
    
    # Later, delete this specific memory
    await memory.delete(memory_id)

asyncio.run(main())

​
Memory Configuration

​
Memory Providers

Marvin uses a memory provider to handle the storage and retrieval of memories. By default, it uses a SQLite-based provider, but you can configure others:

import os

# Configure the memory database location
os.environ["MARVIN_DATABASE_URL"] = "sqlite:///path/to/memories.db"

​
Auto-Use Memory

You can configure a memory to be automatically used before each agent interaction:

import marvin

# Create a memory that's automatically queried
auto_memory = marvin.Memory(
    key="important_facts",
    instructions="Always check these facts before responding",
    auto_use=True
)

# Create an agent with this memory
agent = marvin.Agent(memories=[auto_memory])

# The agent will automatically check the memory before responding
agent.run("Tell me about our company")

​
Best Practices

  • Specific Instructions: Provide clear instructions about what should be stored in each memory
  • Descriptive Keys: Use meaningful keys that describe the memory’s purpose
  • Memory Separation: Create separate memories for different types of information
  • Selective Storage: Don’t store everything - focus on important information that needs to be remembered
  • Regular Maintenance: Periodically review and clean up memories that are no longer needed

​
Advanced Usage

​
Multiple Memory Modules

Agents can have multiple memory modules for different purposes:

import marvin

# Create specialized memories
user_preferences = marvin.Memory(key="user_preferences")
product_knowledge = marvin.Memory(key="product_knowledge")
conversation_history = marvin.Memory(key="past_conversations")

# Create an agent with multiple memories
agent = marvin.Agent(
    name="Customer Support",
    memories=[user_preferences, product_knowledge, conversation_history]
)

​
Memory in Teams

Memory can be shared across a team of agents:

import marvin

# Create a shared memory
project_memory = marvin.Memory(key="project_data")

# Create agents with this memory
agents = [
    marvin.Agent(name="Researcher", memories=[project_memory]),
    marvin.Agent(name="Writer", memories=[project_memory]),
    marvin.Agent(name="Editor", memories=[project_memory])
]

# Create a team with these agents
team = marvin.Swarm(members=agents)

# The team shares knowledge through the memory
team.run("Research quantum computing")
team.run("Write an article about quantum computing")
team.run("Edit the article for clarity")

Memory is a powerful feature that enables more personalized, context-aware AI applications. By effectively using memory, your agents can build up knowledge over time and provide more consistent, informed responses.