ShineMnemos vs Cognee

Cognee and ShineMnemos start from the same premises: memory for agents should be a graph, it should run self-hosted, and it should be reachable over MCP. Cognee (~30k GitHub stars, Apache-2.0) builds that on an ECL pipeline — extract, cognify, load — that ingests documents and other sources and turns them into a queryable knowledge graph. ShineMnemos (Apache-2.0, 316 tests, 11 plugins) does it as a single SQLite file with an in-process node graph and an MCP server. On architecture we are neighbours, not opponents.

The real divergence is what gets validated at write time, and this is worth stating precisely because the two things sound alike. Cognee's `ontology_valid` flag checks whether a fact conforms to the declared schema — the right entity types, the right relation, the right shape. That is genuinely useful: it keeps the graph structurally coherent and makes traversal predictable. But schema conformance says nothing about whether the statement is true. `(ServiceX, deployed_at, 2019-01-01)` can be perfectly ontology-valid and completely wrong.

ShineMnemos validates content. Each fact passes five gates — freshness, source, numbers, consistency, reproducibility — followed by a truth-gate that returns a verdict (up to 6/6) and a confidence score. Facts get a TTL and are auto-invalidated once stale, so the graph degrades on a schedule instead of silently aging. Form validation and truth validation are complementary, not substitutes; we simply chose the harder one to build.

Honest take: where we fall short

• Maturity and adoption. ~30k stars against our early user base. More production hours, more edge cases already found, a larger contributor pool. That gap is not closed by a feature list.
• Ingestion. Cognee's ECL pipeline handles ingesting and structuring bulk source material. ShineMnemos has no equivalent — facts arrive through the gates one at a time. If your job is "turn 4,000 PDFs into a graph," we are the wrong tool today.
• Ontology modelling. We do not offer declarative schemas. If you need a typed, enforced domain ontology, Cognee gives you something we do not, and our gates will not substitute for it.
• Verification is not free. Gates add write latency and reject some true statements. Precision over recall is a deliberate choice, and it is a cost.
• Our latency figure is narrow. ~0.001 ms is an exact in-process lookup over ~10k nodes with no network. It says nothing about end-to-end agent latency.
• 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 results. We are not claiming a number before we have one.

Who should use which

Use Cognee for building a structured knowledge graph out of existing corpora, when ontology enforcement matters and you want a project with real production mileage.

Use ShineMnemos when the question is not "is this well-formed?" but "is the agent allowed to assert this?" — verified facts, confidence and TTL attached, in one local file with nothing leaving the machine.

They compose: Cognee for building the graph, ShineMnemos for the facts your agent states as true.

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