Mem0 v3 Reality Check

Hystersis is not at verified Mem0 v3 parity today. The strongest Hystersis advantage is a self-hosted, queryable Neo4j graph paired with Qdrant and a broad Go API surface. Mem0 v3 is ahead on validated retrieval quality, single-pass extraction, entity linking ergonomics, SDK maturity, and public benchmark evidence. This page is intentionally blunt. Do not market Hystersis as outperforming Mem0 until the benchmark suite in this repository reproduces better numbers on public datasets.

Current Standing

Area	Mem0 v3	Hystersis today	Status
Extraction	Single-pass ADD-only extraction	LLM processing plus async compression; not fully ADD-only	Behind
Entity linking	Built-in entity linking, no graph dependency	Real Neo4j entities and memory links after write	Competitive, more ops-heavy
Retrieval	Semantic + BM25 + entity fusion	Semantic, BM25 adapter, graph spreading activation	Partially competitive
Queryable graph	Removed from OSS v3; entity linking replaces it	Neo4j graph remains queryable	Hystersis advantage
Benchmarks	Published LoCoMo, LongMemEval, BEAM results	Local benchmark runner reports hit rate, MRR, latency, ingest/search/scoring errors, and publishability	Behind, but now measurable
SDKs	Mature Python and TypeScript SDKs	Python/Node SDKs exist; install and endpoint coverage still need hardening	Behind
Async writes	Default async memory processing	Async compression pipeline; memory create still graph-first	Behind
Multi-tenancy	Mature scoping semantics	Hard isolation: auth-bound tenant, spoof rejection, per-tenant Qdrant collections, Tenant CRUD/membership APIs, per-tenant rate limit buckets	Competitive (isolation tests green; membership store is in-process until Neo4j-backed)
Enterprise controls	Platform/enterprise offering	RBAC, audit, SSO code exists	Needs end-to-end product proof

What Hystersis Can Honestly Claim

Self-hosted graph memory with Neo4j and Qdrant.
Queryable graph relationships, not only entity-score boosting.
Memory CRUD, sessions, API keys, RBAC, audit logging, webhooks, skills, sources, and wiki APIs.
Deterministic Qdrant point IDs tied to memory IDs.
Extracted entity materialization into graph nodes and memory-entity links.
Weighted semantic, BM25, and entity rank fusion with signal attribution.
Reproducible benchmark reports that refuse to mark unjudged runs as publishable.
Experimental spreading activation over graph relationships.

What Hystersis Must Not Claim Yet

Better accuracy than Mem0 v3.
97 percent accuracy retention in production.
85-90 percent compression advantage over Mem0.
Proven +23 percent multi-hop improvement.
Full Supermemory-style connector parity.
Fully validated enterprise SSO product flow.

Gap List

Priority	Gap	Required proof
P0	Public benchmark runner	LLM-judged LoCoMo, LongMemEval, and BEAM-style reports committed under `benchmarks/` or `docs/benchmarks/`
P0	ADD-only extraction mode	New facts appended without destructive merge; temporal contradictions preserved
P0	Retrieval fusion	Semantic, BM25, entity, graph, temporal, and reranker signals fused with explainable scores
P0	Entity ingestion invariant	Every memory with extracted entities links to graph entities before vector indexing
P1	SDK parity	Python and Node SDKs pass smoke tests against local and deployed API
P1	Async add mode	API returns quickly while enrichment, compression, and indexing finish in background
P1	Connector parity	Google Drive, Notion, GitHub, files, URLs, and attachments tested through sources APIs
P1	Tenant isolation	Search, get, batch, graph, and source endpoints have cross-tenant negative tests — in progress: auth spoof blocked, search/get/list filter by tenant, per-tenant Qdrant collections, `/tenants` APIs; expand batch/graph/source leakage tests

Recent Core Hardening

The memory write path now persists the primary extracted entity back to the memory node and sends real graph entity IDs into Qdrant metadata. Qdrant no longer defaults entity_id to the memory ID, which keeps memory IDs and graph entity IDs distinct. The benchmark runner now preserves dataset memory IDs during ingestion, reports memory_hit_rate and mrr, and marks results as non-publishable unless every benchmark question is scored by a configured LLM judge with zero ingest/search errors. This is a product-critical invariant for graph-native retrieval:

memory.content
  -> entity extraction
  -> Neo4j Entity nodes
  -> MEMORY_OF links
  -> memory.EntityID / metadata.entity_ids
  -> Qdrant payload memory_id + entity_id + entity_ids
  -> hybrid / spreading retrieval

Benchmark Command

go run ./cmd/benchmark --mock --suite retrieval --dataset locomo

Mock runs are useful for checking benchmark plumbing only. Competitive claims require a live store, an evaluator LLM, and committed benchmark output.

External References

Mem0 Memory Evaluation: https://docs.mem0.ai/core-concepts/memory-evaluation
Mem0 OSS v2 to v3 migration: https://docs.mem0.ai/migration/oss-v2-to-v3
Mem0 Graph Memory: https://docs.mem0.ai/open-source/features/graph-memory

Getting Started

Core Concepts

Integrations

Deployment

Features

Production

Mem0 v3 Reality Check

Mem0 v3 Reality Check

Current Standing

What Hystersis Can Honestly Claim

What Hystersis Must Not Claim Yet

Gap List

Recent Core Hardening

Benchmark Command

External References

​Mem0 v3 Reality Check

​Current Standing

​What Hystersis Can Honestly Claim

​What Hystersis Must Not Claim Yet

​Gap List

​Recent Core Hardening

​Benchmark Command

​External References

Mem0 v3 Reality Check

Current Standing

What Hystersis Can Honestly Claim

What Hystersis Must Not Claim Yet

Gap List

Recent Core Hardening

Benchmark Command

External References