Memgraph CE feature-limits + workarounds#

TL;DR: Memgraph Community Edition silently breaks on 4 features Neo4j users assume work: (1) DDL inside multi-statement transactions, (2) constraints inside transactions, (3) multi-database, (4) MAGE algorithms. The win: native vector-index is FREE in CE 3.9.0 — measured 1ms mean / 2.6ms p95 search latency, 280× speedup over our previous numpy-cosine workaround. Multi-namespace vector-index (3 namespaces × 2829/462/8997 nodes) works out-of-the-box, zero Enterprise license needed.

Origin: Originally written in Hungarian as part of MyForge Vault 11.11 — Superintelligent Vault project. Source: memgraph-ce-feature-limits (Hungarian version).

The Memgraph Community Edition (current build: memgraph/memgraph:latest → 3.9.0) silently rejects or hard-errors on several core features — a production-bug source if you code against it as if it were Neo4j. This doc collects limits actually hit in production and the working workarounds. Update as new limits surface.

What this is NOT#

• NOT a Memgraph-vs-Neo4j shootout — this is a pragmatic CE feature-list, not a benchmark. We chose Memgraph for the in-memory vector-index; your call may differ.
• NOT exhaustive — only limits hit in our production. Other CE limits (HA cluster, audit-log, RBAC) are listed in the migration table but we have NOT independently verified them.
• NOT static — Memgraph ships fast. The "WORKS in CE 3.9.0" callout flipped from earlier docs. Re-verify before relying on it.
• NOT a recommendation against Enterprise — if you need multi-tenancy or HA, pay for Enterprise. The point is CE goes much further than the 2023-era docs suggested.

The 4 main limits#

[!success] What WORKS in CE 3.9.0 (earlier tips are stale) - Native vector index works: CREATE VECTOR INDEX <name> ON :Label(prop) WITH CONFIG {"dimension": 1024, "capacity": 2048, "metric": "cos"} + CALL vector_search.search(...). Free in CE from 3.9.0. The earlier numpy-cosine workaround is no longer needed for new projects (see below). - User management (CREATE USER … IDENTIFIED BY …, DROP USER) — free, but RBAC roles need Enterprise. - Replication basics (SHOW REPLICATION ROLE, REGISTER REPLICA) — work in CE.

Workaround playbook#

1+2. DDL/Constraint in transactions → autocommit + MERGE-by-name#

Forbidden:

with driver.session() as s:
    with s.begin_transaction() as tx:
        tx.run("CREATE INDEX ON :Entity(name)")  # ← BOOM
        tx.run("CREATE CONSTRAINT ON (e:Entity) ASSERT e.name IS UNIQUE")

Correct — two paths:

A) Via mgconsole, row-level autocommit (once, at startup):

docker exec my-memgraph bash -c \
  "echo 'CREATE INDEX ON :Entity(name);' | mgconsole --use_ssl=false"

B) In the Python driver — session.run directly, NOT in a tx (implicit transaction = autocommit):

with driver.session() as s:
    s.run("CREATE INDEX ON :Entity(name)")  # ✅ autocommit

C) App-layer MERGE-dedup instead of constraints:

# No UNIQUE constraint — instead MERGE guarantees one instance per name
tx.run("""
  MERGE (e:Entity {name: $name})
  ON CREATE SET e.created_at = timestamp()
  ON MATCH  SET e.frequency = coalesce(e.frequency, 0) + 1
""", name=ent_name)

3. Multi-database → 1 instance per project + label namespace#

CE has no DATABASE concept — everything lives in the default memgraph DB. Workaround: label-prefix namespacing (:ProjectA_Entity vs :ProjectB_Entity), or a separate Docker container per project on different ports.

4. MAGE algorithms → image swap#

Replace memgraph/memgraph:latest with memgraph/memgraph-mage:latest in docker-compose.yml. Drop-in, same data volume is compatible. Adds ~600MB image size, but you get pagerank / community / node2vec / shortestpath / weakly_connected_components / 50+ algorithms.

Live example — entity-graph extract#

A live entity-graph extracted from a corpus via an LLM pipeline, written to Memgraph CE:

SHOW STORAGE INFO  →  vertex_count: 23964, edge_count: 13812, memory: 105 MiB
SHOW INDEX INFO    →  label+property index :Entity(name) (created via autocommit)

MATCH (n) RETURN labels(n), count(*)
  ["Literal"]   12160
  ["Entity"]     8975
  ["Chunk"]      2829

What we swapped: the original loader started a CREATE CONSTRAINT ... IS UNIQUE inside a transaction → fail. Refactor: push the index migration-style via mgconsole autocommit, and every ingest pass uses MERGE (e:Entity {name: $name}). We get app-layer dedup with index assistance (MERGE looks up via :Entity(name) index → O(log n)).

Migration to Enterprise / MAGE — when#

Native vector-index VERIFIED LIVE#

Memgraph CE 3.9.0 natively supports vector indexes — verified in production.

• Syntax: CREATE VECTOR INDEX vault_chunk_vec ON :Chunk(embedding) WITH CONFIG {"dimension": 1024, "metric": "cos", "capacity": 8192};
• Query: CALL vector_search.search('vault_chunk_vec', limit, query_vector) YIELD node, similarity
• Benchmark: pure search mean 1ms / p95 2.6ms (280× speedup vs numpy-cosine workaround)
• Top-K result agreement with numpy: 100% (3 queries × k=3 byte-exact)

Multi-namespace vector-index confirmation#

One Memgraph DB can host multiple independent label-prop vector indexes in parallel — 0 mucking-around (SHOW INDEX INFO lists them together). Example:

label+property_vector  Chunk       embedding  2829   (vault content)
label+property_vector  SkillChunk  embedding   462   (skill pool)
label+property_vector  Entity      embedding  8997   (entity-graph)

• 3 different namespaces (content / skill-pool / entity-graph)
• 0 cross-namespace interference (separate index, separate cosine space)
• Latency independent: search 8-13ms on the 462-skill namespace, mean 1ms on the 2829-content namespace
• CREATE VECTOR INDEX <name> ON :<Label>(<prop>) WITH CONFIG ... is one CQL statement per namespace

Reusable insight: Memgraph CE multi-tenant vector-search out-of-the-box (no Neo4j-style license requirement), namespace per :<Label> trivially separated.

Audio overview#

• EN narration (Charon voice): [[.vault-nb/audio-overviews/memgraph-ce-feature-limits.en.mp3]]
• HU narration (Kore voice): [[.vault-nb/audio-overviews/memgraph-ce-feature-limits.hu.mp3]]

Generated via Gemini 3.1 Flash TTS preview. ~1-2 minutes each. See gemini-3-1-flash-tts-pipeline for the pipeline.

Related#

• mgclient-autocommit-silent-rollback — the related Python driver pitfall
• verification-step-before-claim — verify vendor features before implementing workarounds