langchain-hypergraph

LangChain chat memory and vector store backed by the Temporal Hypergraph Engine.

pip install langchain-hypergraph
docker run -d -p 50051:50051 ghcr.io/divyansh151005/temporal-hypergraph-server

For development from a clone: pip install -e integrations/python-client -e integrations/langchain.

from langchain_hypergraph import TemporalHypergraph

memory = TemporalHypergraph(api_key="your_key")  # a TemporalHypergraphMemory on session "default"
memory = TemporalHypergraph(api_key="your_key", session_id="user-42")

The server address comes from endpoint= or $HYPERGRAPH_ENDPOINT (default localhost:50051).

Chat memory

TemporalHypergraphMemory is a BaseChatMessageHistory, so it plugs into RunnableWithMessageHistory:

from langchain_core.runnables.history import RunnableWithMessageHistory

chat = RunnableWithMessageHistory(
    prompt | llm,
    lambda session_id: TemporalHypergraphMemory(session_id=session_id),
    input_messages_key="input",
    history_messages_key="history",
)

• Each message becomes one hyperedge connecting the session:<id> node, a per-message record node, and any ids in additional_kwargs["entities"].
• messages returns the session oldest first (history_limit= keeps the newest N).
• relevant_context(query, k, k_hops=2) returns the most relevant facts, also reaching those linked through the messages' entities. Pass embeddings= for similarity ranking.
• ToolMessage.tool_call_id and other message fields besides the text and entities are not stored.

Vector store

from langchain_hypergraph import TemporalHypergraphVectorStore

store = TemporalHypergraphVectorStore(embeddings)
store.add_texts(
    ["Acme acquired Beta"],
    [{"entities": ["org:acme", "org:beta"], "valid_from": 1696118400, "event_type": "ACQUISITION"}],
)
store.similarity_search("who owns Beta?", k=4, as_of=1700000000)
store.similarity_search("Beta CEO", seed_entities=["org:acme"], k_hops=3)

• Recognized metadata: entities, valid_from, valid_to (epoch seconds), event_type, confidence. Other metadata is dropped; the engine stores only the text and the graph structure.
• Ids are assigned by the engine and returned by add_texts.
• Search keywords: as_of, session_id, seed_entities, k_hops, decay_lambda.
• The embedding dimension must match the server's --embedding-dim.

Tests

cargo build --release -p grpc-server   # for the live tests; they skip otherwise
pytest integrations/langchain