langchain-memoryrouter

Persistent, user-scoped memory for LangChain and LangGraph agents, backed by MemoryRouter.

MemoryRouter gives each user in an AI product a private memory vault. This package connects that memory layer to LangChain through tools, LangGraph nodes, and a BaseStore implementation.

• Documentation
• API reference
• Get a Memory Key
• Security
• Public SDK repository

Install

pip install langchain-memoryrouter

Python 3.9–3.13 is supported. The package requires langchain-core>=0.3.0, langgraph>=0.2.0, and httpx>=0.25.

Create a Memory Key at app.memoryrouter.ai. In multi-user products, map each app user to a separate Memory Key so each user has a separate vault.

Choose an integration pattern

Pattern 1: tools

Works with LangChain bind_tools() and LangGraph create_react_agent().

from langchain_memoryrouter import create_memory_tools

retain, recall = create_memory_tools(memory_key="mk_user_123")
model_with_tools = model.bind_tools([retain, recall])

The package exposes two tool primitives:

• memoryrouter_retain stores sanitized conversation text with /v1/memory/ingest
• memoryrouter_recall searches memory with /v1/memory/search

MemoryRouter is retain and recall; this package does not add a separate reflect tool.

Pattern 2: LangGraph nodes

Use nodes when you want memory to run automatically in the graph.

from langchain_memoryrouter import create_recall_node, create_retain_node

recall_node = create_recall_node(memory_key_config_key="memory_key")
retain_node = create_retain_node(memory_key_config_key="memory_key")

result = graph.invoke(
    {"messages": messages},
    config={"configurable": {"memory_key": "mk_user_123", "thread_id": "chat_abc"}},
)

create_recall_node calls /v1/memory/prepare and returns a ready-to-inject memory context block. create_retain_node sanitizes conversation messages before calling /v1/memory/ingest.

Pattern 3: BaseStore

Use MemoryRouterStore when a LangGraph integration expects a LangChain BaseStore.

from langchain_memoryrouter import MemoryRouterStore

store = MemoryRouterStore(memory_key="mk_user_123")
graph = builder.compile(store=store)

MemoryRouter is semantic memory, not a literal key-value database. mset retains text into the user's vault. mget searches by key and returns the best matching memory content. The public MemoryRouter API does not expose arbitrary LangChain store-key listing or deletion, so yield_keys returns an empty iterator and mdelete is a documented no-op.

Storage hygiene

Every write path sanitizes messages before ingest. The integration stores conversation text only:

• Keeps HumanMessage text as user
• Keeps AIMessage text as assistant
• Drops SystemMessage and ToolMessage content
• Drops tool calls, tool-call IDs, and function-call arguments
• Drops AI messages that contain only tool calls and no text

This behavior is enforced in tools, nodes, and store writes; it is not a configuration option.

Never commit a Memory Key, paste one into an issue, or share private memory content in a support request. See the MemoryRouter security page and this package's SECURITY.md.

API targeted

Base URL: https://api.memoryrouter.ai

Auth: Authorization: Bearer mk_xxx

Retain:

POST /v1/memory/ingest
{
  "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
  "session_id": "optional",
  "model": "optional",
  "embeddings": "optional"
}

Recall:

POST /v1/memory/search
{"query": "what does the user prefer", "limit": 10}

Graph recall context:

POST /v1/memory/prepare
{
  "messages": [{"role": "user", "content": "..."}],
  "session_id": "optional",
  "density": "default",
  "context_limit": 10
}

For the current API surface, use the live API reference.

Support and package history

• Product and account help: hello@memoryrouter.ai
• Package changes: CHANGELOG.md
• Security reporting: SECURITY.md
• MemoryRouter company profile: LinkedIn

The package is maintained by MemoryRouter and available under the MIT License.