@engramviz/adapter-langgraph

Evidence-preserving instrumentation for the LangGraph long-term Store interface.

The adapter wraps InMemoryStore, PostgresStore, MongoDBStore, and other BaseStore-compatible implementations without importing a concrete store. It captures put, search, get, delete, and direct batch calls inside the active Engram turn.

import { InMemoryStore } from "@langchain/langgraph";
import {
  instrumentLangGraphStore,
  langGraphMemoryIds
} from "@engramviz/adapter-langgraph";
import { EngramClient } from "@engramviz/sdk";

const engram = new EngramClient({ adapter: "langgraph" });
const store = instrumentLangGraphStore(new InMemoryStore(), engram);

await engram.withTurn(turnOptions, async (turn) => {
  const results = await store.search(["users", userId, "memories"], {
    query: "Where does the user live?",
    limit: 3
  });

  const loadedIds = langGraphMemoryIds(results);
  await turn.load(loadedIds);
  return graph.invoke(input, { context: { userId } });
});

Search results are captured as retrieval candidates. They are not assumed to have reached model context; call turn.load(...) only for memories actually included in model input.

LangGraph checkpointers are intentionally outside this adapter. Checkpoints are thread execution state, while LangGraph Store values are the durable cross-thread memory boundary Engram observes.