LangGraph

Install the adapter for LangGraph 1.2:

pip install "protoprompt[langgraph]"

Scope-safe Store

ProtoPromptStoreAdapter is a complete sync/async LangGraph BaseStore. It preserves logical namespaces for graph code while adding an opaque physical prefix derived from a host-owned MemoryScope:

from langgraph.store.memory import InMemoryStore
from protoprompt import MemoryScope
from protoprompt.integrations import ProtoPromptStoreAdapter

store = ProtoPromptStoreAdapter(
    InMemoryStore(),
    scope=MemoryScope(tenant="acme", user="alice"),
)
store.put(("memories",), "contract", {"renewal": "May"})

The boundary applies to get, put, delete, search, namespace listing, batch operations, and all async counterparts. Use one adapter instance per trusted scope. Do not build MemoryScope from model-generated tool arguments.

Ready context node

Use the async factory with graph.ainvoke or the sync factory with graph.invoke:

from protoprompt.integrations import create_build_context_node

graph.add_node(
    "build_context",
    create_build_context_node(builder, chat_id="case-42"),
)

The node reads state["query"], falling back to the newest text message. It writes context and content-free context_provenance. It never rewrites the messages field, so it is safe with LangGraph message reducers. Customize ContextInput with input_factory= when profiles or per-graph policy need to be supplied by trusted application code.

State ownership

Keep the two lifecycles distinct:

• a LangGraph checkpointer owns per-thread execution state and message history;
• a scoped LangGraph Store or a protoprompt backend owns cross-thread memory.

The offline example stores one profile across two threads while each thread's history remains independent:

python examples/langgraph_memory.py