Compare

Genesys vs Cognee

Cognee and Genesys are the two most alike systems in this comparison — both graph-native, both with an explicit memory-lifecycle stage, both rejecting memory-as-flat-vector-store. Cognee's strength is the ingestion-to-graph pipeline: 38+ formats, a structured ECL pipeline, 14 retrieval modes. Genesys's is the legibility of the retention decision: causal edges, a multiplicative score you can audit force by force, and a certified recall number. This is the fairest architectural fight in the set — here is where each one wins.

The steelman

What Cognee is.

Cognee is the graph-native, pipeline-first memory platform, and of everyone in this list it is the closest in spirit to Genesys — which is exactly why the comparison has to be careful and fair. It is open source (Apache-2.0), self-hostable for free, and built around an ECL pipeline — Extract, Cognify, Load — that turns raw data into a queryable knowledge graph. The Cognify stage is a real, six-step transformation: classify documents, check permissions, extract chunks, use an LLM to pull entities and relationships, generate summaries, then embed everything and commit edges to the graph. It ingests 38+ formats (PDF, CSV, JSON, audio, images, code) and exposes 14 retrieval modes, from classic RAG to chain-of-thought graph traversal, so agents can reason over structure rather than just semantic surface.

The most Genesys-adjacent piece is Cognee's memify stage: after ingestion, it refines the graph by pruning stale nodes, strengthening frequent connections, reweighting edges based on usage signals, and adding derived facts. That is a genuine memory-lifecycle idea, not just storage — and we should say so plainly. Cognee has real traction (a $7.5M seed led by Pebblebed in February 2026, live in 70+ companies, pipeline volume past a million runs in its last reported year). If you want a graph-vector hybrid you fully control, with rich ingestion and many retrieval strategies, Cognee is a serious, credible choice.

Architecture

Knowledge graph vs. causal graph.

Cognee reweights an entity/relationship graph by usage; Genesys walks a causal graph and scores retention as a single multiplicative decay. Same rejection of the flat vector store; different semantics for what an edge means and how a memory earns its place.

DimensionCogneeGenesys
Build modelECL pipeline (Extract → Cognify → Load) over 38+ formatsPer-turn store; heavy work (embedding, causal inference) is async
GraphEntity/relationship knowledge graph (graph-vector hybrid)One causal graph per user; typed causal edges
Edge meaningOntology / entity relationships, usage-reweightedcaused_by, supersedes, context — causal semantics the scorer walks
Lifecyclememify: prune stale nodes, strengthen frequent edges, reweight by usage, derive factsStatus FSM (TAGGED → ACTIVE → EPISODIC → DORMANT) + consolidation + reactivation cascade
Retention ruleReweighting + pruning (usage-driven)Conjunctive prune: decay-score < 0.01 and orphaned and unpinned and not core
ScoringUsage signals reweight edgesMultiplicative: relevance × causal connectivity × reactivation frequency
ExplainabilityGraph is inspectable; 14 retrieval modes incl. CoT traversalmemory_explain returns causal chain + live per-force score; memory_traverse returns edges

Both systems reject “memory = flat vector store,” both are graph-native, and both have an explicit lifecycle stage — this is the fairest architectural fight in the set. The differences are in emphasis and semantics. Cognee's edges are primarily ontological (entities and their relationships), reweighted by usage; Genesys's edges are primarily causal (why a memory exists, what it superseded, what context it belongs to), and its retention is a single multiplicative decay formula with a conjunctiveforgetting rule — a node survives unless it is simultaneously irrelevant, orphaned, unpinned, and non-core. Cognee's memify strengthens and prunes by usage frequency; Genesys separates the three forces (relevance, connectivity, reactivation) so you can inspect which force is keeping a memory alive.

Put bluntly: if you want the richest ingestion pipeline and the most retrieval strategies over a self-hosted graph, Cognee is ahead. If you want the memory decision— what's recalled, what's forgotten, and why — to be causal and auditable to the individual force, that is the Genesys axis. Where Cognee's strength is the ingestion-to-graph pipeline (38+ formats, 14 retrieval modes), Genesys's is the legibility of the retention decision — memory_explain's score model and memory_amend's supersede-don't-overwrite semantics.

Benchmark reality

The number, measured.

One yardstick, applied to everyone the same way — and a blank cell where we could not measure honestly, rather than a fabricated one.

Genesys, under the frozen LoCoMo protocol evaluated July 2026, scores 85.55 ± 0.37 J-score — the mean ± standard deviation over ten independent full runs (84.87, 85.84, 85.71, 85.58, 85.26, 85.78, 85.52, 86.17, 85.52, 85.19). The protocol is frozen and identical for every scored system: n = 1,540 questions, categories 1–4, answerer and judge both gpt-4o-mini at temperature 0.

We could not source a Cognee LoCoMo result under a comparable configuration, so this page does not assign Cognee a number. Cognee publishes its own benchmark material, but we found nothing under this exact frozen judge/answerer/dataset setup that we can cite honestly, and we will not convert an unlike figure into a false comparison. For context, the two competitor numbers we can cite under the shared protocol, next to ours:

SystemLoCoMo J (frozen protocol)
Mem066.9
Zep75.14
Genesys85.55 ± 0.37

Comparability rule. A LoCoMo number is comparable only when the dataset, the answering model, and the judge all match. The rows above share all three. Vendor self-reports produced under different answerer/judge configurations are not comparable and are excluded here — and we hold ourselves to exactly the same rule: our own non-comparable run, which changes only the frozen answerer, is disclosed below and never quoted as the headline. See the LoCoMo citation surface for the exact claim and the ten runs, and the methodology dossier for the full ledger, including the negative results.

Churn5.4% of answers regenerate differently on temperature-0 re-runs; we certify with a ten-run mean, not a single lucky run. The ± 0.37 spread is the honest measure of that noise.
The trade+170 gained / −72 lost vs. baseline (a 2.4 : 1 trade), flagged RED by our own persistence gate and published. Certified as net-better with a disclosed trade, not as strictly better.
The ceilingAn oracle probe caps this protocol at 94.9. Our non-comparable 89.68 row (modern answerer, same memories and judge) changes the frozen answerer, so it stays a footnote — never the headline.

If Cognee publishes a LoCoMo number under this configuration, we will add it verbatim. We would rather show a blank cell than a fabricated one.

The verdict

Which to choose.

Neither system is a strict superset of the other. Choose on the axis your workload actually lives on — ingestion breadth and retrieval variety, or causal recall you can audit and cite.

Choose Cognee
IngestionYour problem is ingestion-to-graph breadth: 38+ formats and a structured ECL pipeline that turns raw data into a knowledge graph for you.
RetrievalYou want many retrieval strategies out of the box — 14 modes, from classic RAG to chain-of-thought graph traversal — to experiment with.
Self-hostYou want a fully self-hostable, Apache-2.0 graph-vector platform, free forever locally, with usage-based cloud ($2.50/1M tokens, $5 per additional workspace) if you don't self-host.
ReweightingYou want usage-driven graph refinement — memify: reweight frequent edges, prune stale nodes, derive facts — and that model fits how your data gets used.
Track recordYou're standing up a knowledge-graph memory for a data-heavy domain (research, enterprise documents) and value Cognee's production track record.
Choose Genesys
Causal recallCausal recall matters more than ontological breadth. You need to traverse why a memory exists and what it superseded — causal edges, not just entity relationships. memory_explain returns the chain and a live per-force score.
AuditableYou want the retention decision auditable to the force — relevance, connectivity, and reactivation kept separate — and forgetting that requires all conjunctive criteria, not a single usage heuristic.
SupersedeCorrections should supersede, not silently reweight. memory_amend records a new node that supersedes the old and keeps the history queryable.
CertifiedYou want a certified, checkable recall number — 85.55 ± 0.37 under a frozen protocol with an open harness, including the failures and the measured ceiling.
PortableYou want one portable, user-owned graph across ChatGPT, Claude, and any MCP client — inspectable and theirs.
Sources · retrieved 2026-07-19

Where these facts come from.

PricingFree $0 (1M tokens / 1 workspace) · Standard $2.50/1M tokens + $5 per additional workspace · Enterprise custom; self-host free forever, cloud on gpt-oss-120b. cognee.ai/pricing
EngineApache-2.0; ECL pipeline; graph-vector hybrid. github.com/topoteretes/cognee
PipelineECL, memify, and the 14 retrieval modes, described by Cognee. cognee.ai/blog · how-cognee-builds-ai-memory
Funding$7.5M seed (Cognee announcement). cognee.ai/blog · seed round
TractionPebblebed-led seed, 70+ companies, pipeline runs past 1M in its last reported year. techinberlin.com · Cognee seed
GenesysThe certified LoCoMo figures — exact claim, ten runs, comparison, disclosures. genesys.astrixlabs.ai/benchmarks/locomo
Competitor facts retrieved 2026-07-19 via web search. Pricing and usage rates change — re-verify against the primary sources before relying on them. Genesys numbers are the certified frozen-protocol figures at genesys.astrixlabs.ai/benchmarks/locomo.