I kept hitting disk bloat and python GIL bottlenecks when using standard VectorDBs for continuous langgraph loops. so i gutted the db approach and built a headless rust daemon (null-drift).
it manages the agent’s memory as a continuous array using geometric decay. memory stays completely flat at O(1) no matter how long your graph runs.
i just wrote the python wrapper so it inherits directly from BaseStore. it maps your langgraph namespaces directly to isolated phase spaces natively using aput and aget.
from nulldrift_agents import NullDriftLangGraphStore
from langgraph.graph import StateGraph
store = NullDriftLangGraphStore(gateway_url="http://127.0.0.1:8000")
repo is here if anyone building local-first infra wants to test it out and try to break the async rust daemon! null-drift