bitwikiorg """Middleware for loading agent memory/context from AGENTS.md files.
This module implements support for the AGENTS.md specification (https://agents.md/), loading memory/context from configurable sources and injecting into the system prompt.
Overview
AGENTS.md files provide project-specific context and instructions to help AI agents work effectively. Unlike skills (which are on-demand workflows), memory is always loaded and provides persistent context.
Usage
from deepagents import MemoryMiddleware
from deepagents.backends.filesystem import FilesystemBackend
# Security: FilesystemBackend allows reading/writing from the entire filesystem.
# Either ensure the agent is running within a sandbox OR add human-in-the-loop (HIL)
# approval to file operations.
backend = FilesystemBackend(root_dir="/")
middleware = MemoryMiddleware(
backend=backend,
sources=[
"~/.deepagents/AGENTS.md",
"./.deepagents/AGENTS.md",
],
)
agent = create_deep_agent(middleware=[middleware])
Memory Sources
Sources are simply paths to AGENTS.md files that are loaded in order and combined. Multiple sources are concatenated in order, with all content included. Later sources appear after earlier ones in the combined prompt.
File Format
AGENTS.md files are standard Markdown with no required structure. Common sections include:
- Project overview
- Build/test commands
- Code style guidelines
- Architecture notes """
from future import annotations
import logging from collections.abc import Awaitable, Callable from typing import TYPE_CHECKING, Annotated, NotRequired, TypedDict
from langchain_core.runnables import RunnableConfig
if TYPE_CHECKING: from deepagents.backends.protocol import BACKEND_TYPES, BackendProtocol
from langchain.agents.middleware.types import ( AgentMiddleware, AgentState, ModelRequest, ModelResponse, PrivateStateAttr, ) from langchain.tools import ToolRuntime from langgraph.runtime import Runtime
from deepagents.middleware._utils import append_to_system_message
logger = logging.getLogger(name)
class MemoryState(AgentState):
"""State schema for MemoryMiddleware.
Attributes:
memory_contents: Dict mapping source paths to their loaded content.
Marked as private so it's not included in the final agent state.
"""
memory_contents: NotRequired[Annotated[dict[str, str], PrivateStateAttr]]
class MemoryStateUpdate(TypedDict):
"""State update for MemoryMiddleware."""
memory_contents: dict[str, str]
MEMORY_SYSTEM_PROMPT = """ {agent_memory}
**Learning from feedback:**
- One of your MAIN PRIORITIES is to learn from your interactions with the user. These learnings can be implicit or explicit. This means that in the future, you will remember this important information.
- When you need to remember something, updating memory must be your FIRST, IMMEDIATE action - before responding to the user, before calling other tools, before doing anything else. Just update memory immediately.
- When user says something is better/worse, capture WHY and encode it as a pattern.
- Each correction is a chance to improve permanently - don't just fix the immediate issue, update your instructions.
- A great opportunity to update your memories is when the user interrupts a tool call and provides feedback. You should update your memories immediately before revising the tool call.
- Look for the underlying principle behind corrections, not just the specific mistake.
- The user might not explicitly ask you to remember something, but if they provide information that is useful for future use, you should update your memories immediately.
**Asking for information:**
- If you lack context to perform an action (e.g. send a Slack DM, requires a user ID/email) you should explicitly ask the user for this information.
- It is preferred for you to ask for information, don't assume anything that you do not know!
- When the user provides information that is useful for future use, you should update your memories immediately.
**When to update memories:**
- When the user explicitly asks you to remember something (e.g., "remember my email", "save this preference")
- When the user describes your role or how you should behave (e.g., "you are a web researcher", "always do X")
- When the user gives feedback on your work - capture what was wrong and how to improve
- When the user provides information required for tool use (e.g., slack channel ID, email addresses)
- When the user provides context useful for future tasks, such as how to use tools, or which actions to take in a particular situation
- When you discover new patterns or preferences (coding styles, conventions, workflows)
**When to NOT update memories:**
- When the information is temporary or transient (e.g., "I'm running late", "I'm on my phone right now")
- When the information is a one-time task request (e.g., "Find me a recipe", "What's 25 * 4?")
- When the information is a simple question that doesn't reveal lasting preferences (e.g., "What day is it?", "Can you explain X?")
- When the information is an acknowledgment or small talk (e.g., "Sounds good!", "Hello", "Thanks for that")
- When the information is stale or irrelevant in future conversations
- Never store API keys, access tokens, passwords, or any other credentials in any file, memory, or system prompt.
- If the user asks where to put API keys or provides an API key, do NOT echo or save it.
**Examples:**
Example 1 (remembering user information):
User: Can you connect to my google account?
Agent: Sure, I'll connect to your google account, what's your google account email?
User: john@example.com
Agent: Let me save this to my memory.
Tool Call: edit_file(...) -> remembers that the user's google account email is john@example.com
Example 2 (remembering implicit user preferences):
User: Can you write me an example for creating a deep agent in LangChain?
Agent: Sure, I'll write you an example for creating a deep agent in LangChain <example code in Python>
User: Can you do this in JavaScript
Agent: Let me save this to my memory.
Tool Call: edit_file(...) -> remembers that the user prefers to get LangChaincode examples in JavaScript
Agent: Sure, here is the JavaScript example<example code in JavaScript>
Example 3 (do not remember transient information):
User: I'm going to play basketball tonight so I will be offline for a few hours.
Agent: Okay I'll add a black to your calendar.
Tool Call: create_calendar_event(...) -> just calls a tool, does not commit anything to memory, as it is transient information
class MemoryMiddleware(AgentMiddleware):
"""Middleware for loading agent memory from AGENTS.md files.
Loads memory content from configured sources and injects into the system prompt.
Supports multiple sources that are combined together.
Args:
backend: Backend instance or factory function for file operations.
sources: List of `MemorySource` configurations specifying paths and names.
"""
state_schema = MemoryState
def __init__(
self,
*,
backend: BACKEND_TYPES,
sources: list[str],
) -> None:
"""Initialize the memory middleware.
Args:
backend: Backend instance or factory function that takes runtime
and returns a backend. Use a factory for StateBackend.
sources: List of memory file paths to load (e.g., `["~/.deepagents/AGENTS.md",
"./.deepagents/AGENTS.md"]`).
Display names are automatically derived from the paths.
Sources are loaded in order.
"""
self._backend = backend
self.sources = sources
def _get_backend(self, state: MemoryState, runtime: Runtime, config: RunnableConfig) -> BackendProtocol:
"""Resolve backend from instance or factory.
Args:
state: Current agent state.
runtime: Runtime context for factory functions.
config: Runnable config to pass to backend factory.
Returns:
Resolved backend instance.
"""
if callable(self._backend):
# Construct an artificial tool runtime to resolve backend factory
tool_runtime = ToolRuntime(
state=state,
context=runtime.context,
stream_writer=runtime.stream_writer,
store=runtime.store,
config=config,
tool_call_id=None,
)
return self._backend(tool_runtime)
return self._backend
def _format_agent_memory(self, contents: dict[str, str]) -> str:
"""Format memory with locations and contents paired together.
Args:
contents: Dict mapping source paths to content.
Returns:
Formatted string with location+content pairs wrapped in <agent_memory> tags.
"""
if not contents:
return MEMORY_SYSTEM_PROMPT.format(agent_memory="(No memory loaded)")
sections = []
for path in self.sources:
if contents.get(path):
sections.append(f"{path}\n{contents[path]}")
if not sections:
return MEMORY_SYSTEM_PROMPT.format(agent_memory="(No memory loaded)")
memory_body = "\n\n".join(sections)
return MEMORY_SYSTEM_PROMPT.format(agent_memory=memory_body)
async def _load_memory_from_backend(
self,
backend: BackendProtocol,
path: str,
) -> str | None:
"""Load memory content from a backend path.
Args:
backend: Backend to load from.
path: Path to the AGENTS.md file.
Returns:
File content if found, None otherwise.
"""
results = await backend.adownload_files([path])
# Should get exactly one response for one path
if len(results) != 1:
raise AssertionError(f"Expected 1 response for path {path}, got {len(results)}")
response = results[0]
if response.error is not None:
# For now, memory files are treated as optional. file_not_found is expected
# and we skip silently to allow graceful degradation.
if response.error == "file_not_found":
return None
# Other errors should be raised
raise ValueError(f"Failed to download {path}: {response.error}")
if response.content is not None:
return response.content.decode("utf-8")
return None
def _load_memory_from_backend_sync(
self,
backend: BackendProtocol,
path: str,
) -> str | None:
"""Load memory content from a backend path synchronously.
Args:
backend: Backend to load from.
path: Path to the AGENTS.md file.
Returns:
File content if found, None otherwise.
"""
results = backend.download_files([path])
# Should get exactly one response for one path
if len(results) != 1:
raise AssertionError(f"Expected 1 response for path {path}, got {len(results)}")
response = results[0]
if response.error is not None:
# For now, memory files are treated as optional. file_not_found is expected
# and we skip silently to allow graceful degradation.
if response.error == "file_not_found":
return None
# Other errors should be raised
raise ValueError(f"Failed to download {path}: {response.error}")
if response.content is not None:
return response.content.decode("utf-8")
return None
def before_agent(self, state: MemoryState, runtime: Runtime, config: RunnableConfig) -> MemoryStateUpdate | None:
"""Load memory content before agent execution (synchronous).
Loads memory from all configured sources and stores in state.
Only loads if not already present in state.
Args:
state: Current agent state.
runtime: Runtime context.
config: Runnable config.
Returns:
State update with memory_contents populated.
"""
# Skip if already loaded
if "memory_contents" in state:
return None
backend = self._get_backend(state, runtime, config)
contents: dict[str, str] = {}
for path in self.sources:
content = self._load_memory_from_backend_sync(backend, path)
if content:
contents[path] = content
logger.debug(f"Loaded memory from: {path}")
return MemoryStateUpdate(memory_contents=contents)
async def abefore_agent(self, state: MemoryState, runtime: Runtime, config: RunnableConfig) -> MemoryStateUpdate | None:
"""Load memory content before agent execution.
Loads memory from all configured sources and stores in state.
Only loads if not already present in state.
Args:
state: Current agent state.
runtime: Runtime context.
config: Runnable config.
Returns:
State update with memory_contents populated.
"""
# Skip if already loaded
if "memory_contents" in state:
return None
backend = self._get_backend(state, runtime, config)
contents: dict[str, str] = {}
for path in self.sources:
content = await self._load_memory_from_backend(backend, path)
if content:
contents[path] = content
logger.debug(f"Loaded memory from: {path}")
return MemoryStateUpdate(memory_contents=contents)
def modify_request(self, request: ModelRequest) -> ModelRequest:
"""Inject memory content into the system message.
Args:
request: Model request to modify.
Returns:
Modified request with memory injected into system message.
"""
contents = request.state.get("memory_contents", {})
agent_memory = self._format_agent_memory(contents)
new_system_message = append_to_system_message(request.system_message, agent_memory)
return request.override(system_message=new_system_message)
def wrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], ModelResponse],
) -> ModelResponse:
"""Wrap model call to inject memory into system prompt.
Args:
request: Model request being processed.
handler: Handler function to call with modified request.
Returns:
Model response from handler.
"""
modified_request = self.modify_request(request)
return handler(modified_request)
async def awrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
) -> ModelResponse:
"""Async wrap model call to inject memory into system prompt.
Args:
request: Model request being processed.
handler: Async handler function to call with modified request.
Returns:
Model response from handler.
"""
modified_request = self.modify_request(request)
return await handler(modified_request)