Compare / vs Cognee
Source-anchored evidence graphvs institutional knowledge graph.
Cognee is a real knowledge-graph engine, and for teams that need an institutional KG with deep ETL it is the better answer. Zephr is the evidence graph above the editor: every claim carries a receipt, every retrieval proves authorization, and the system is allowed to refuse to answer. The right pick depends on whether your bottleneck is graph schema or graph honesty.
Different substrates, different guarantees.
Cognee optimizes for an institutional knowledge graph with rich ETL. Zephr optimizes for a source-anchored evidence graph whose edges carry receipts. The columns are not a scorecard; they are different answers to different questions — pick the one whose failure you are more worried about.
| Feature | Cognee | Zephr |
|---|---|---|
| Substrate | Generic knowledge graph (entities, relations, ontologies) | Evidence graph (claims, sources, receipts, review state) |
| Memory scope | Project / tenant (one global graph) | Per-worktree, per-session, per-project — scope-bound |
| Cross-client | SDK + REST API; integrations layered on the graph | MCP-native — works across all connected agents out of the box |
| Pipeline | ECL pipeline (extract, cognify, memify); rich ETL primitives | Append-only evidence graph; consolidation is reviewable, not automatic |
| Honest abstention | Graph may return low-confidence results; first-class abstention is up to the caller | First-class abstention — the kit renders "we do not know" rather than a guess |
| Trust governance | Per-tenant ACL; authorization at the query layer | Trust Firewall — authorization before retrieval, signed by a human, audit-trailed |
| Open-core | Yes (Apache-2.0 / MIT) | Yes — local SQLite, no account required |
Public positioning snapshot, re-checked August 2026 — not live data. Cognee is actively developed and may add comparable capabilities; this is a dated view, not a forecast.
Checkable capabilities, including where Zephr is partial.
A matrix that came out all-green for Zephr would mean the rows were chosen to flatter us. Institutional knowledge graph is a real Cognee win; source-anchored evidence, first-class abstention, and signed review are the Zephr bets.
| Substrate | Institutional KG | Source-anchored evidence | First-class abstention | Signed review |
|---|---|---|---|---|
| Cognee | ●supported | ◐partial | ◐partial | ◐partial |
| Zephrthis page | ◐partial | ●supported | ●supported | ●supported |
Substrate lens · public positioning re-checked 2026-08-22 · ADR-Z-16 · packages/core/src/memory/* · packages/mcp-server/src/transport.ts
verdicts are testable and dated — no scores, no winners
Cognee is the better knowledge graph for an enterprise stack.
Zephr is not better at everything, and pretending otherwise would undermine the one thing this product is for. Cognee's strengths below are real, and for teams that need a real institutional knowledge graph they are the deciding factor.
Institutional knowledge graph
A real entity-relationship substrate with ontologies, rules, and a query language that can answer "what does our org know about X" with the right schema in place.
ECL pipeline depth
Extract, cognify, memify are first-class primitives. Teams that need a real ETL over their memory get a real ETL, with the schema, the rules, and the operators they would build by hand otherwise.
Open-source engine
Self-hostable, inspectable, and the substrate is the product. If you wanted a graph database with memory semantics, Cognee is the closer answer than rolling your own.
Entity modelling
A proper entity layer is the part that makes a memory system composable with the rest of an enterprise stack — Cognee treats that as the primary design surface, not a side feature.
Public positioning re-checked August 2026. Cognee is a moving target; treat any single feature as a dated snapshot rather than a permanent ranking.
Where Zephr adds value.
Zephr is not a faster Cognee. It is a different substrate — source-anchored evidence, first-class abstention, and trust governance — and it is worth more the more you need the graph to be auditable rather than merely queryable.
Source-anchored evidence
A claim is stored with the files, commits, and line ranges it rests on. The graph is the load-bearing part — if a node says it knows something, the receipt is right there on the edge.
Honesty lattice
First-class abstention as a UI state, an API state, and a vocabulary. A memory system that guesses when it does not know is a memory system that lies, and the kit is the visible proof that Zephr does not.
Cross-tool portability
Memory is held above the editor, so it survives the moment you move from Claude Code to Cursor to a custom MCP client. The graph belongs to the workflow, not to one tool.
Human-signed review
Confirmation is recorded with a human identity and a `confirmation_audit` row. A graph that auto-confirms its own beliefs is a graph with no review, not a graph with reviews.
Choose Cognee or Zephr?
The choice comes down to which failure you are more worried about: a knowledge graph whose schema is wrong, or a memory system that confidently answers when it should abstain.
Neither failure mode is exotic: a generic knowledge graph that returns confidently wrong answers is a live risk every week, and a source-anchored evidence graph with the wrong schema is a first-week cost. Which one you are absorbing decides the category.
Choose Cognee if…
You need an institutional knowledge graph with rich ETL, entity-relationship modelling, and a query layer your existing stack already speaks — and you can accept that the system will return results even when it should not.
Choose Zephr if…
You work across multiple coding agents, need per-worktree memory scope, and want every claim to carry a receipt, every retrieval to prove authorization, and the system to refuse to answer when it cannot.
See how Zephr enforces trust.
The trust model covers authorization before retrieval, scope binding, and evidence receipts — the three things this comparison keeps returning to.