Why mnestic

If you are evaluating an embedded database for agent memory in 2026, the first question is not a feature question. It is: will this engine still be maintained in two years? Two of the three serious embedded graph engines for this niche have already answered it the wrong way.

The field

• CozoDB — the original, and still a remarkable design: Datalog graph queries, HNSW vector search, and full-text search in one embeddable engine. Its last release (v0.7.6) shipped in 2023; the last commit on main landed 2024-12-04. The community's own issue tracker asked "is this still maintained?" and got no answer.
• Kùzu — the most credible embedded graph database of its generation, archived October 2025 when its team was acqui-hired. Community forks carry it with varying activity; the most active is pivoting toward lakehouse analytics.
• mnestic — a hard fork of CozoDB, carried forward under its own name the way LadybugDB carried Kùzu and libSQL carried SQLite. Full upstream history preserved, original remote kept as upstream, MPL-2.0 headers intact.

The stewardship record

Maintenance is a claim you verify, not a promise you accept. The record since the fork:

What the fork has added

Everything upstream had, plus — each with its own page:

• Bitemporality — valid time and transaction time on facts, :as_of queries, ::history, and audited eviction (::evict): deletion with a trail, never silent.
• Hybrid retrieval — one call fusing an HNSW vector leg, a BM25 full-text leg, and a k-hop graph leg, with RRF/MMR and per-leg score explanations. Since 0.13.0 the graph leg can run as a budgeted cheapest-first expansion.
• Cached graph projections and interruptible queries with budgets — the operational tooling long-lived memory needs.
• Provenance semirings, skyline aggregates, non-blocking HNSW builds, planner fixes (equality pushdown, join reorder, factorized counting), corruption-repair primitives, and ULIDs.

The full inventory is in What mnestic adds and the release history.

Coming from CozoDB

mnestic is a drop-in continuation: the library name is still cozo, your CozoScript runs unchanged, and adopting the fork also hands you ~30 upstream commits that were merged after v0.7.6 but never released — including a real silent-wrong-answer fix. See Migrating from CozoDB.

Coming from Kùzu

Honesty first: mnestic is not a Kùzu drop-in. The query language is Datalog (a read-only Cypher surface exists but is alpha), and on fixed-shape analytical join benchmarks the columnar Kùzu lineage genuinely wins — we say so plainly and don't compete there.

What mnestic offers instead, for the agent-memory workload specifically:

• Indexes that see your writes. Agent memory is written continuously, not batch-loaded. mnestic maintains FTS and vector indexes transactionally on every write; there is no "rows added after index creation aren't indexed" failure mode, and ::reindex exists as an explicit repair primitive.
• Recursion at scale. Recursive reachability at LDBC-SNB sf1 runs in 100–161 ms flat at any depth in mnestic, where the leading Kùzu fork — measured in its best idiomatic query form — exceeds 120 s from depth 3.
• Bitemporal facts and audited eviction, native to the engine.
• A maintained mainline rather than a fork ecosystem still finding its footing.

Use it from the stack you already run

You don't have to adopt Datalog to adopt the engine: pip install mnestic-mcp (any MCP client, including Claude Code and Cursor), langgraph-store-mnestic (LangGraph BaseStore), langchain-mnestic and llama-index-vector-stores-mnestic (vector stores with the graph legs exposed) — and via LangChain's provider interface, Mem0 OSS runs on mnestic unchanged.

mnestic is alpha and single-node, and says so. What it will not be is abandoned quietly: the changelog, the CI, and the cadence are the proof you can check.