langchain-sqlite-graph-memory

A LangChain retriever over your own markdown notes, with the links between them. It runs sqlite-graph-memory: vector search finds the entry points, the [[wikilinks]] you already wrote are the graph, and a cross-encoder reranks what the 1-hop expansion pulled in. Everything runs on your machine: no server, no graph database, no LLM call at retrieval time.

pip install "langchain-sqlite-graph-memory[lite]"   # ONNX models, starts in seconds
sgm-index /path/to/your/notes                      # builds ./index once

from langchain_sqlite_graph_memory import SQLiteGraphMemoryRetriever

retriever = SQLiteGraphMemoryRetriever(index_dir="./index", k=8)

for doc in retriever.invoke("why did we drop the graph database?"):
    print(doc.metadata["rank"], doc.metadata["title"], doc.metadata["via_graph"])

Each result is a Document. page_content holds the matched chunk of the note. metadata holds path, title, date, score (cross-encoder), rank, and via_graph. via_graph is True when the note came in only through a wikilink from a top vector hit, so you can tell associative recall from direct recall.

Options

In an agent

from langchain_core.tools import create_retriever_tool

recall = create_retriever_tool(
    retriever, "recall_notes",
    "Search my notes and the notes they link to. Use for anything I wrote down before.",
)

Scope, stated plainly

• This adapter calls the core library and adds nothing to retrieval. It calls the same functions sgm-ask calls (query_sims, retrieve_candidates, expand_1hop, rerank_candidates). A test pins this: the adapter's order must equal a direct call of those functions. On a real vault of 29,474 notes (105,383 chunks), the adapter returned the same 12 notes in the same order as sgm-ask --graph.
• The index is two files on disk, an embedding matrix (.npy) and note metadata (.pkl), written by sgm-index. Wikilinks are read from the notes at query time. SQLite stores the core library's session memory and A/B logs. The retriever needs only the two index files.
• Embeddings: intfloat/multilingual-e5-base. Reranker: cross-encoder/mmarco-mMiniLMv2-L12-H384-v1. Both work for many languages.
• Not implemented yet: a VectorStore (adding documents through LangChain) and a LangGraph long-term memory store. Indexing goes through sgm-index.

Tests

pip install -e ".[test]"
pytest

The suite includes LangChain's standard retriever tests (langchain-tests). It runs with fake models on a four-note vault: no download, no GPU, a few seconds.

License

MIT