mnestic 0.11.1

Embedded graph + vector + full-text database with Datalog — a maintained fork of CozoDB, tuned as a substrate for agentic memory.

mnestic (Python)

Embedded graph + vector + full-text database with Datalog queries — a maintained fork of CozoDB, tuned as a substrate for agentic memory. This package is the in-process Python binding (no server required).

mnestic is not the official CozoDB and is not affiliated with or endorsed by its original authors. All credit for the original design belongs to Ziyang Hu and the Cozo Project Authors. See the fork repository for provenance and licensing.

pip install mnestic

from mnestic import CozoDbPy

db = CozoDbPy("mem", "", "{}")  # engines: "mem", "sqlite" (file path), "rocksdb" (dir path)
db.run_script("?[x] <- [[1],[2],[3]]", {}, False)

# One-call hybrid retrieval (HNSW + full-text fused with Reciprocal Rank Fusion),
# over a relation that has an HNSW index and an FTS index:
hits = db.hybrid_search({
    "relation": "docs",
    "vector_index": "vec", "query_vector": [0.1, 0.9], "vector_k": 5,
    "fts_index": "fts", "query_text": "vector search", "fts_k": 5,
})
# -> {"headers": ["id", "score"], "rows": [["d3", 0.033], ...], "next": None}

# Pass "detailed": True for per-leg contributions — one row per (item, leg)
# with the within-leg rank the fusion used and the leg's raw score:
# headers ["id","score","list_id","leg_rank","leg_score"]

The "rocksdb" persistent backend now ships in the published wheel — CozoDbPy("rocksdb", "./my.db", "{}") works straight from pip install mnestic. The source distribution stays SQLite/compact-only, so the persistent engine is wheel-only.

Upgrade note (0.10.6): a persistent database whose relation catalogs were last written by a build older than 0.10.0 could fail to open with Cannot deserialize relation metadata from bytes after upgrading to 0.10.0–0.10.5. 0.10.6 fixes this — legacy catalogs open again with no migration, so upgrade to 0.10.6 if you carry a pre-0.10.0 database.

run_script takes an optional timeout= — a per-query wall-clock budget in seconds; on expiry the query raises an eval::timeout error. db.set_default_query_timeout(secs) sets a Db-wide default and db.default_query_timeout() reads it back; the effective budget for a query is the minimum of that default and any per-call timeout.

New in 0.11.1: skyline aggregates, no binding code. pareto_min and pareto_max keep, per group, the Pareto frontier of a numeric vector — the points no other point dominates — and are ordinary CozoScript aggregates, so they run straight from the wheel through run_script with nothing to register: db.run_script("?[f] := offer[p, q], v = [p, -q], f = pareto_min(v)", {}, False). pareto_min treats smaller as better and pareto_max larger; mix objectives (minimize price, maximize quality) by negating the maximized components, as above. A malformed operand (non-list, non-numeric component, NaN, or empty vector) raises. It is a non-breaking, purely additive patch — no existing query changes result.

For idiomatic LangChain / LlamaIndex usage, install the integration packages (langchain-mnestic, llama-index-vector-stores-mnestic).

The query language (CozoScript / Datalog) and engine semantics follow CozoDB; see the upstream documentation and the fork changelog.

License

Mozilla Public License 2.0. Original work © 2022 The Cozo Project Authors; fork modifications © 2026 Shan Rizvi.