langgraph-minnsdb

LangGraph long-term memory on MinnsDB, the temporal memory database for agents.

LangGraph's built-in stores keep the latest value of each item. When a user moves from London to Berlin, London is gone, and the agent can't answer "where did they live in March?". MinnsDB keeps every version with the time it held, so you can.

Two pieces:

• MinnsDBStore: a drop-in LangGraph BaseStore. Same put / get / search / list_namespaces as any store, plus history() and get_as_of().
• MinnsDBMemory: conversation memory in MinnsDB's temporal graph. Send it chat turns and MinnsDB extracts facts, superseding old ones rather than piling them up. Ask what's true about someone now or on any date. Comes as agent tools.

Install

pip install langgraph-minnsdb

You need a MinnsDB server. To run one locally:

git clone https://github.com/Minns-ai/MinnsDB
cd MinnsDB && docker compose up --build

It listens on http://localhost:3000. Point the package at it with MINNSDB_URL, and MINNSDB_API_KEY if the server has auth turned on.

The store

from langgraph_minnsdb import MinnsDBStore

store = MinnsDBStore()                 # MINNSDB_URL, or http://localhost:3000
graph = builder.compile(store=store)   # use it like any LangGraph store

Inside a node it behaves like InMemoryStore or PostgresStore:

store.put(("users", "priya"), "home", {"city": "London"})
store.put(("users", "priya"), "home", {"city": "Berlin"})

store.get(("users", "priya"), "home").value             # {'city': 'Berlin'}
store.search(("users",), filter={"city": "Berlin"})
store.list_namespaces(prefix=("users",))

What only this store can do:

store.get_as_of(("users", "priya"), "home", some_time_before_the_move).value
# {'city': 'London'}

for v in store.history(("users", "priya"), "home"):
    print(v.value, v.valid_from, v.valid_until)
# {'city': 'London'} <when it was written> <when Berlin replaced it>
# {'city': 'Berlin'} <when it was written> None

get_as_of takes a timezone-aware datetime. Deleting an item closes its last version, and history() still shows it. The async methods (aput, aget, asearch, ...) work too. examples/time_travel.py runs the whole thing inside a LangGraph graph.

Items live in one MinnsDB table (langgraph_store by default, created on first use; pass table= to change it).

Graph memory for agents

from langgraph_minnsdb import MinnsDBMemory

memory = MinnsDBMemory(case_id="user-priya")

memory.remember([{"role": "user", "content": "I'm Priya. I live in London and work at Monzo."}])
# months later
memory.remember([{"role": "user", "content": "I've moved to Berlin and started at Stripe."}])

memory.facts_about("Priya")                     # what holds now
memory.facts_about("Priya", at="2026-03-01")    # what held on that date
memory.facts_about("Priya", history=True)       # every version, with valid_from and valid_until
memory.recall("Where does Priya work?")["answer"]

remember also takes LangChain messages, so you can pass state["messages"] after each turn. System prompts and tool messages are skipped, and the agent's own replies are ignored unless you set include_assistant_facts=True. It waits for extraction to finish; pass wait=False to get a background job instead and check it with memory.job(job_id). If extraction finds no facts in what the user said, remember raises a NoFactsExtracted warning. The usual cause is a MinnsDB server with no LLM key.

Give an agent the tools, for example with LangChain's create_agent:

from langchain.agents import create_agent

agent = create_agent("anthropic:claude-sonnet-5-5", tools=memory.as_tools())

The server-side LLM is MinnsDB's own (LLM_API_KEY in its environment), not your agent's model.

For anything else, run MinnsQL directly:

memory.query('MATCH (a)-[r]->(b) WHEN "2026-03-01" WHERE a.name = "Priya" RETURN type(r), b.name')

Limits

• search(query=...) ranks by the share of query words found in each item. It does not use embeddings. For meaning-based recall, use MinnsDBMemory.
• search and list_namespaces read the whole table. Fine for thousands of items; past 100,000 rows MinnsDB stops paging and the store raises an error rather than return part of the table.
• TTL is not supported.
• MinnsDBMemory writes and reads under its case_id, so one user's facts never answer for another. Raw query() is not scoped.

Run the tests

pip install -e ".[test]"
pytest                                          # unit tests
MINNSDB_URL=http://localhost:3000 pytest        # plus integration tests against your server
MINNSDB_URL=... MINNSDB_LLM=1 pytest            # plus fact extraction, if the server has an LLM key

Licence

MIT. MinnsDB itself is AGPL-3.0; this package only talks to it over HTTP.