langchain-mimir

Persistent, local-first, encrypted memory for LangChain, backed by Mimir — an open-source (MIT) memory engine with FTS5 + dense hybrid search and optional AES-256-GCM encryption, exposed over the Model Context Protocol (MCP) stdio transport.

It gives a LangChain agent durable memory that survives across runs and processes, stored in a single local SQLite file you control — no external service, no cloud.

What you get

This package wraps Mimir using the modern langchain-core interfaces:

• create_mimir_tools(client) — a pair of StructuredTools (mimir_remember / mimir_recall) you give to an agent so it can manage its own long-term memory via tool calls. This is the current-recommended LangChain pattern (the legacy Memory / ConversationBufferMemory classes are deprecated).
• MimirRetriever — a BaseRetriever returning Documents, for drop-in use in RAG chains and anywhere LangChain accepts a retriever (.invoke(query)).
• MimirClient — the low-level MCP stdio client, if you want direct access.

Prerequisite: the mimir binary

This package talks to a local mimir executable via JSON-RPC over stdio. You must have it installed:

• Download a release from https://github.com/Perseus-Computing-LLC/mimir/releases, or build from source (cargo build --release), and put mimir on your $PATH.
• Or pass an absolute path: MimirClient(mimir_binary="/path/to/mimir").

On Windows the binary may be named mimir.exe; ensure its directory is on PATH, or pass the full path.

Install

pip install langchain-mimir

Usage

As agent tools

from langchain_mimir import MimirClient, create_mimir_tools

client = MimirClient(db_path="~/.langchain/mimir.db")
tools = create_mimir_tools(client)  # [mimir_remember, mimir_recall]

# Bind to any tool-calling model / agent:
from langchain.chat_models import init_chat_model

llm = init_chat_model("anthropic:claude-sonnet-4-5")
llm_with_memory = llm.bind_tools(tools)

resp = llm_with_memory.invoke("Remember that my favorite language is Rust.")
# ... the model will call mimir_remember; execute the tool call as usual.

As a retriever

from langchain_mimir import MimirClient, MimirRetriever

client = MimirClient(db_path="~/.langchain/mimir.db")
client.remember("The capital of France is Paris.")

retriever = MimirRetriever(client=client, k=5)
docs = retriever.invoke("What is the capital of France?")
print(docs[0].page_content)  # -> "The capital of France is Paris."

Direct client

from langchain_mimir import MimirClient

client = MimirClient(db_path="~/.langchain/mimir.db")
client.remember("Project deadline is July 15.", tags=["project", "deadline"])
items = client.recall("when is the deadline")
print(items[0]["text"])

How it works

MimirClient spawns mimir --db <path> as a subprocess and speaks JSON-RPC 2.0 (MCP) over its stdin/stdout. A background reader thread and a lock make calls thread-safe and timeout-bounded. Memories are stored via mimir_remember and retrieved via mimir_recall.

License

MIT © 2026 Perseus Computing LLC