MindGraph vs Mem0 vs Zep
All three give your agent memory. The honest differences: MindGraph puts a source on every fact, tracks contradictions over time, and connects your live SQL database — fusing your rows with what your documents say about them. Here’s where each tool leads.
The most-adopted memory API — strong defaults, broad framework integrations, the easy on-ramp. Provenance isn’t its focus.
The strongest temporal story — a bi-temporal knowledge graph with edge validity. The closest tool to us on time and lineage.
Ships the cognition, not an empty engine — and fuses your structured systems with your documents under one typed, sourced contract.
We benchmark in the open (LongMemEval, reproducible harness) and refuse self-scored leaderboards. This table reflects each tool’s emphasis as of mid-2026 — corrections welcome.
| Mem0 | Zep | MindGraph | |
|---|---|---|---|
| Core approach | Extracted text memories | Temporal knowledge graph | Typed cognitive graph + extraction |
| Source / provenance on every fact | Shallow (name field) | Source filtering | Claim → evidence → source |
| Contradiction & supersession tracking | Edge validity | Typed Contradicts / Supersedes | |
| Temporal model (as-of / history) | Bi-temporal | Validity / as-of | |
| Connect a live SQL / Postgres database | One-way JSON ingest | Auto-mapped to typed objects | |
| Fuse structured rows + unstructured docs on one object | By identity resolution | ||
| Typed domain ontology (shipped, not BYO) | 6 layers, drafted from your schema | ||
| Human-in-the-loop curation / approval | Review inboxes, nothing top-down | ||
| Where the learning lives | Vendor store | Graph (Neo4j) | Your graph, model-portable |
| MCP server + TS / Python SDKs | |||
| Open-source engine | Library | Graphiti | Mnestic (Rust, MPL-2.0) |
Which should you pick?
if you want the fastest on-ramp to conversational memory and native integrations with CrewAI / LangGraph, and per-fact lineage isn’t a requirement yet.
if a temporal knowledge graph at scale is the core need and your memory is built from conversational/JSON sources rather than a live operational database.
if your agent reasons over a real operational database and unstructured context, and a wrong or unsourced answer has consequences — legal, finance, ops, research. You want provenance on every fact, contradictions surfaced for a human to settle, and the expertise to stay yours when you swap models.
Coming from Mem0 or Zep?
The concepts map cleanly. You add memory, you retrieve it — MindGraph just returns typed, sourced graph context alongside the text.
| Mem0 / Zep concept | MindGraph equivalent |
|---|---|
| add(messages, user_id) | POST /ingest/document → extracts entities, claims & sources |
| search(query, user_id) | POST /retrieve { action: "hybrid" } → chunks + typed graph |
| graph episodes / facts | Claims linked to Evidence → Source (provenance built in) |
| user / agent scoping | per-agent identity + private/shared layers |
| — (no equivalent) | Connect a Postgres DB → typed objects, fused with docs |
Keep your existing source of truth — MindGraph indexes a read-only projection; it never writes back to your database.
Run both in parallel during migration. Point one agent at MindGraph and compare answers — no rip-and-replace.
The MCP server drops into Claude / Cursor / your framework, so the integration surface is familiar.
See it on your own data
Connect a read-only Postgres database and ask your agent something only the fusion can answer. Free to start.
Building something where trust matters? We work closely with a few design partners — get in touch.