Memory

Store and retrieve facts with a vector database

A Memory gives an agent tools to add, search, and delete facts. It uses a vector-database provider, separately from the database that stores thread history.

The default provider, chroma-ephemeral, keeps data in memory. It does not survive a process restart. Install Chroma separately and choose chroma-db for persistence:

uv add chromadb

import marvin

preferences = marvin.Memory(
    key="user_preferences",
    provider="chroma-db",
    instructions="Store the user's stated writing preferences.",
)
assistant = marvin.Agent(memories=[preferences])
assistant.run("Remember that I prefer short answers.")

chroma-db stores data under ~/.marvin/memory/chroma by default. Reuse the same provider configuration and memory key to access stored facts later.

Direct operations

import asyncio
import marvin

async def main():
    memory = marvin.Memory(key="writing_preferences", provider="chroma-db")
    memory_id = await memory.add("The user prefers short answers.")
    results = await memory.search("Preferred answer length", n=1)
    await memory.delete(memory_id)
    print(results)

asyncio.run(main())

Search returns a dictionary of memory IDs and their contents. A memory key may contain letters, numbers, and underscores.

Retrieval and providers

By default, the agent chooses whether to use memory tools. auto_use=True also queries memory before the agent runs, using recent messages.

You can pass a MemoryProvider instance or a supported provider string: chroma-ephemeral, chroma-db, chroma-cloud, lancedb, postgres, or qdrant. Each provider needs its own dependencies and configuration; see the memory API.

Memory can be shared by multiple agents. Use separate keys for facts that should be searched separately. MARVIN_DATABASE_URL configures thread history, not vector memory.