MemoryMiddleware | deepagents | LangChain Reference

LangChain Reference home page

GitHub
Main Docs

Overview
Agent
Middleware
Backends
Sandboxes
Subagents

Deep Agents Code

⌘I

Chat LangChain

What can I help with?

Enter to send·Shift+Enter new line

MemoryMiddleware

Copy

MemoryMiddleware(
  self,
  *,
  backend: BackendProtocol,
  sources: list[str],
  add_cache_control:

Bases

AgentMiddleware[MemoryState, ContextT, ResponseT]

Constructors

Attributes

Methods

View source on GitHub

bool

False

system_prompt

str

None

MEMORY_SYSTEM_PROMPT

)

Parameters

Name	Type	Description
`backend`*	`BackendProtocol`	Backend instance for reading memory sources.
`sources`*	`list[str]`	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.
`add_cache_control`	`bool`	Default:`False` If `True`, tag the last system-message content block with `cache_control: {"type": "ephemeral"}` when the request model is `ChatAnthropic`. This creates a second prompt-cache breakpoint that pairs with `AnthropicPromptCachingMiddleware`'s breakpoint on the static system prompt, keeping the memory block boundary cached across turns (memory content would otherwise shift after every update and invalidate the prefix cache). No-ops on non-Anthropic models; Bedrock and Vertex wrappers do not qualify.
`system_prompt`	`str \| None`	Default:`MEMORY_SYSTEM_PROMPT`

constructor

__init__

Name	Type
backend	BackendProtocol
sources	list[str]
add_cache_control	bool
system_prompt	str \| None

attribute

trace_policy

Omit hook inputs from traces by default; set a TracePolicy to override.

attribute

state_schema: MemoryState

attribute

sources: sources

attribute

system_prompt: system_prompt

method

before_agent

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.

method

abefore_agent

Load memory content before agent execution.

Loads memory from all configured sources and stores in state. Only loads if not already present in state.

method

modify_request

Inject memory content into the system message.

method

wrap_model_call

Wrap model call to inject memory into system prompt.

method

awrap_model_call

Async wrap model call to inject memory into system prompt.

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. See constructor for the full argument list.

System-prompt fragment template. Must contain a {agent_memory} slot for runtime memory substitution. Pass None to skip appending entirely (memory is still loaded into state["memory_contents"]).

Sunbelt Computer Software

PL/B Language Development and Support

Chat LangChain

What can I help with?

Menu

MemoryMiddleware

Bases

Constructors

Attributes

Methods

Parameters