LangGraph

ChronoVec can give a LangGraph execution an isolated semantic-memory branch. This is useful when a graph explores several candidate plans, policies, or agent trajectories and should publish only the selected result.

LangGraphMemory itself has no dependency on the langgraph package -- it only maps thread ids to branches, framework-agnostic by design. The optional extra is for actually building and running a real LangGraph graph (as in the example below), not for importing the adapter:

pip install "chronovec[langgraph]"

The LangGraphMemory adapter maps one LangGraph thread_id to one ChronoVec branch:

from chronovec import AgentMemory
from chronovec.integrations import LangGraphMemory

memory = AgentMemory(dimensions=384)
branches = LangGraphMemory(memory)

candidate = branches.open("candidate-a")
candidate.add("new-fact", embedding, text="candidate fact")

# Publish only after the graph evaluator accepts the candidate.
branches.merge("candidate-a")
# Or: branches.discard("candidate-a")

open() is idempotent for an existing thread, which matters when LangGraph replays a node from a checkpoint. It also reattaches to an active same-named branch restored by AgentMemory.load, so a workflow can resume after a process restart when both the graph checkpoint and ChronoVec checkpoint are durable. A failed run can call cleanup(thread_id); close() discards every branch still owned by the adapter. LangGraph’s checkpointer stores graph state, while ChronoVec provides snapshot-isolated semantic state and atomic merge/discard.

The runnable example is examples/langgraph_branching_memory.py. It evaluates narrow, correct, and unsafe refund-policy candidates, then merges only the accepted policy into the main memory. It uses a deterministic local embedding, so no model key or external service is required.