Memory

LLMs work best when they remember what has already been said. Thinagents ships with pluggable memory back-ends so you can choose between fast in-process storage, file-based persistence, or full SQL durability.

Available Memory Stores

Quick Start (In-Memory)

from thinagents import Agent
from thinagents.memory import InMemoryStore

agent = Agent(
    name="Memory Demo",
    model="openai/gpt-4o-mini",
    memory=InMemoryStore(),  # volatile, but ultra-fast
)

A Short Conversation

conv_id = "demo-1"

print(await agent.arun("Hi, I'm Alice!", conversation_id=conv_id))
# → "Hello Alice, nice to meet you."

print(await agent.arun("What is my name?", conversation_id=conv_id))
# → "Your name is Alice."

Switching to Persistent Stores

Need resilience across restarts? Swap the store—no other changes.

from thinagents.memory import FileMemory, SQLiteMemory

file_agent = Agent(
    name="File Mem Agent",
    model="openai/gpt-4o-mini",
    memory=FileMemory(storage_dir="./agent_mem"),
)

db_agent = Agent(
    name="SQLite Mem Agent",
    model="openai/gpt-4o-mini",
    memory=SQLiteMemory(db_path="./agent_mem.db"),
)

Because the memory layer is pluggable, you can even implement your own store (Redis, DynamoDB, etc.) by subclassing BaseMemory.

Tips

1. Choose InMemoryStore for unit tests or short-lived processes.
2. Prefer FileMemory or SQLiteMemory when you need persistence but want a zero-dependency setup.
3. Clear memory per user or conversation by simply changing conversation_id.

Storing Tool Artifacts in Memory

One advantage of InMemoryStore is that it can also store tool artifacts (such as files, datasets, or other outputs from tools) in memory. To enable this, pass store_tool_artifacts=True:

agent = Agent(
    name="Memory Demo",
    model="openai/gpt-4o-mini",
    memory=InMemoryStore(store_tool_artifacts=True),
)

Now, any artifacts returned by tools will be kept in memory and available for later steps in the same conversation.