ShineMnemos vs Zep / Graphiti

Graphiti (~30k GitHub stars, Apache-2.0) is the engine behind Zep, and its central idea is bitemporality: every edge carries `valid_at` and `invalid_at`, so the graph records both when a fact was true in the world and when the system learned it. That is the most rigorous time model in agent memory today. If you need to reconstruct what the system believed at a given moment — audit trails, post-incident review, "why did the agent say that last Tuesday" — Graphiti has an answer we do not.

ShineMnemos attacks a different axis. Instead of recording every belief with its validity interval, we try to keep bad facts out of the graph in the first place. Each write passes five gates — freshness, source, numbers, consistency, reproducibility — plus a truth-gate returning a verdict (up to 6/6) and a confidence score. Facts carry a TTL and are auto-invalidated once stale. Bitemporality answers *when was this believed true*; our gates answer *should this have been written at all*. Both are legitimate; they are not the same guarantee, and neither replaces the other.

The other difference is operational surface. Zep's Community Edition — the self-hostable server — was deprecated in 2025. Self-hosting today means running Graphiti against Neo4j or FalkorDB: several moving systems to deploy, secure, back up, and upgrade. ShineMnemos is one SQLite file, a Python process, and an MCP server. You copy the file to back it up and delete the file to forget everything. Nothing leaves the machine.

Honest take: where we fall short

• No time travel. This is the big one. We tell you what is true now and drop what is not. We cannot reconstruct the belief state of three weeks ago. If your compliance or debugging story requires that, Graphiti is the correct choice and we are not a substitute.
• Scale ceiling. Neo4j and FalkorDB scale far past a single SQLite file. Our design assumes a memory that fits comfortably on one machine, because that is the deployment we optimise for.
• Managed option. Zep has a hosted product with commercial support. We have a repository, a licence, and an MCP endpoint.
• Maturity. ~30k stars and years of production deployments against our early adoption. More real-world failure modes have been found in their code than in ours.
• Gates have a cost. Verification adds write latency and will occasionally reject a true statement. Precision over recall is intentional, and it is a trade.
• Our latency figure is narrow. ~0.001 ms is an exact in-process lookup over ~10k nodes with no network involved — not an end-to-end agent latency claim.
• No published benchmark yet. HaluMem (MemTensor) reports 17–30% hallucination rates in QA across existing memory systems. We are running ShineMnemos through HaluMem now and will publish the numbers. Until they exist, take no number from us.

Who should use which

Use Zep / Graphiti when the history of your knowledge is itself the product: audit trails, point-in-time queries, large graphs, or a managed service with support.

Use ShineMnemos when you want fewer wrong facts rather than a complete record of them, when the deployment must stay on one machine, and when a single file you can copy or delete beats a multi-service graph stack.

Try ShineMnemos → — Apache-2.0, one SQLite file, MCP server, pay-per-call in USDC. No signup.