Deep Agents Memory

Short-term (StateBackend): Persists within a single thread, lost when thread ends Long-term (StoreBackend): Persists across threads and sessions Hybrid (CompositeBackend): Route different paths to different backends

FilesystemMiddleware provides tools: ls, read_file, write_file, edit_file, glob, grep

The read_file tool natively supports image files (.png, .jpg, .jpeg, .gif, .webp) across all backends, returning them as multimodal content blocks.

agent = create_deep_agent() # Default: StateBackend result = agent.invoke({ "messages": [{"role": "user", "content": "Write notes to /draft.txt"}] }, config={"configurable": {"thread_id": "thread-1"}})

/draft.txt is lost when thread ends

</python>
<typescript>
Default StateBackend stores files ephemerally within a thread.
```typescript
import { createDeepAgent } from "deepagents";

const agent = await createDeepAgent();  // Default: StateBackend
const result = await agent.invoke({
  messages: [{ role: "user", content: "Write notes to /draft.txt" }]
}, { configurable: { thread_id: "thread-1" } });
// /draft.txt is lost when thread ends

store = InMemoryStore()

composite_backend = lambda rt: CompositeBackend( default=StateBackend(rt), routes={"/memories/": StoreBackend(rt)} )

agent = create_deep_agent(backend=composite_backend, store=store)

/draft.txt -> ephemeral (StateBackend)

/memories/user-prefs.txt -> persistent (StoreBackend)

</python>
<typescript>
Configure CompositeBackend to route paths to different storage backends.
```typescript
import { createDeepAgent, CompositeBackend, StateBackend, StoreBackend } from "deepagents";
import { InMemoryStore } from "@langchain/langgraph";

const store = new InMemoryStore();

const agent = await createDeepAgent({
  backend: (config) => new CompositeBackend(
    new StateBackend(config),
    { "/memories/": new StoreBackend(config) }
  ),
  store
});

// /draft.txt -> ephemeral (StateBackend)
// /memories/user-prefs.txt -> persistent (StoreBackend)

config2 = {"configurable": {"thread_id": "thread-2"}} agent.invoke({"messages": [{"role": "user", "content": "Read /memories/style.txt"}]}, config=config2)

Thread 2 can read file saved by Thread 1

</python>
<typescript>
Files in /memories/ persist across threads via StoreBackend routing.
```typescript
// Using CompositeBackend from previous example
const config1 = { configurable: { thread_id: "thread-1" } };
await agent.invoke({ messages: [{ role: "user", content: "Save to /memories/style.txt" }] }, config1);

const config2 = { configurable: { thread_id: "thread-2" } };
await agent.invoke({ messages: [{ role: "user", content: "Read /memories/style.txt" }] }, config2);
// Thread 2 can read file saved by Thread 1

agent = create_deep_agent( backend=FilesystemBackend(root_dir=".", virtual_mode=True), # Restrict access interrupt_on={"write_file": True, "edit_file": True}, checkpointer=MemorySaver() )

Agent can read/write actual files on disk

</python>
<typescript>
Use FilesystemBackend for local development with real disk access and human-in-the-loop.
```typescript
import { createDeepAgent, FilesystemBackend } from "deepagents";
import { MemorySaver } from "@langchain/langgraph";

const agent = await createDeepAgent({
  backend: new FilesystemBackend({ rootDir: ".", virtualMode: true }),
  interruptOn: { write_file: true, edit_file: true },
  checkpointer: new MemorySaver()
});

Security: Never use FilesystemBackend in web servers - use StateBackend or sandbox instead.

agent = create_deep_agent( backend=LocalShellBackend(root_dir=".", env={"PATH": "/usr/bin:/bin"}) )

Security risks: commands run directly on host with no sandboxing. Use Sandbox backends for production.
</python>
</ex-localshell-backend>

<ex-store-in-custom-tools>
<python>
Access the store directly in custom tools for long-term memory operations.
```python
from langchain.tools import tool, ToolRuntime
from langchain.agents import create_agent
from langgraph.store.memory import InMemoryStore

@tool
def get_user_preference(key: str, runtime: ToolRuntime) -> str:
    """Get a user preference from long-term storage."""
    store = runtime.store
    result = store.get(("user_prefs",), key)
    return str(result.value) if result else "Not found"

@tool
def save_user_preference(key: str, value: str, runtime: ToolRuntime) -> str:
    """Save a user preference to long-term storage."""
    store = runtime.store
    store.put(("user_prefs",), key, {"value": value})
    return f"Saved {key}={value}"

store = InMemoryStore()

agent = create_agent(
    model="gpt-4.1",
    tools=[get_user_preference, save_user_preference],
    store=store
)

• Backend type and configuration
• Routing rules for CompositeBackend
• Root directory for FilesystemBackend
• Human-in-the-loop for file operations

What Agents CANNOT Configure

• Tool names (ls, read_file, write_file, edit_file, glob, grep)
• Access files outside virtual_mode restrictions
• Cross-thread file access without proper backend setup
• LocalShellBackend: virtual_mode provides no security with shell access (commands can reach any path)

CORRECT

agent = create_deep_agent(backend=lambda rt: StoreBackend(rt), store=InMemoryStore())

</python>
<typescript>
StoreBackend requires a store instance.
```typescript
// WRONG
const agent = await createDeepAgent({ backend: (c) => new StoreBackend(c) });

// CORRECT
const agent = await createDeepAgent({ backend: (c) => new StoreBackend(c), store: new InMemoryStore() });