langgraph-store-mnestic

A LangGraph BaseStore backed by mnestic — an embedded graph + vector + full-text database (a maintained fork of CozoDB). One process, one file, no server: agent long-term memory with atomic batches and in-engine hybrid retrieval (BM25 + vector fused with Reciprocal Rank Fusion — a fusion that Postgres + pgvector, LangGraph's default production store, cannot do in one system).

mnestic is a maintained fork of CozoDB; it is not the official CozoDB. Original design credit belongs to Ziyang Hu and the Cozo Project Authors.

pip install langgraph-store-mnestic

from langgraph_store_mnestic import MnesticStore

store = MnesticStore(
    engine="sqlite", path="memory.db",          # or engine="mem" for ephemeral
    index={"dims": 1536, "embed": my_embed_fn}, # any LangGraph IndexConfig embed
    ttl={"default_ttl": 60 * 24, "refresh_on_read": True},  # optional
)

store.put(("users", "u1"), "pref", {"text": "prefers window seats"})
hits = store.search(("users", "u1"), query="seating preference")

Use it in a graph like any store:

graph = builder.compile(store=store)

What you get

• Atomic batch() — every write in a batch lands in one engine transaction. Under LangGraph's parallel fan-out this is the difference between correct recall and silently losing reads (a non-atomic prototype lost 36% of concurrent semantic reads; this store's concurrency suite gates on zero).
• Hybrid search(query=...) — BM25 keyword + HNSW vector legs, namespace filter pushed into both index searches, RRF-fused and hydrated in a single engine snapshot. Without an index config the store is BM25-only (keyword search still works, no embeddings needed).
• Collision-safe namespaces — any label content stays distinct (('a.b',) can never collide with ('a', 'b')).
• TTL — per-item or default_ttl (minutes), lazy expiry on every read path, optional refresh-on-read, and a start_ttl_sweeper() background reclaimer.
• Sync + async — full BaseStore surface; a* methods run the sync engine calls on the default executor.

LangGraph Platform (langgraph.json)

Custom stores are loadable by import path (alpha upstream feature):

# src/agent/store.py
from contextlib import asynccontextmanager
from langgraph_store_mnestic import MnesticStore

@asynccontextmanager
async def generate_store():
    store = MnesticStore(engine="sqlite", path="memory.db", index={...})
    try:
        yield store
    finally:
        store.close()

{ "store": { "path": "./src/agent/store.py:generate_store" } }

Notes

• Requires Python ≥ 3.10 and langgraph-checkpoint >= 4.1.
• The wheel build of mnestic ships mem, sqlite, and rocksdb engines (the sdist has no rocksdb).
• The text index applies English stemming + stopwords: queries made only of very common words match nothing on the keyword leg (the vector leg still answers when an index config is present).
• Filter operators supported in search(filter=...): $eq, $ne, $gt, $gte, $lt, $lte, with InMemoryStore-parity semantics.

License

Mozilla Public License 2.0.