Genesys vs Supermemory
Both give an AI agent persistent memory, but they optimize different ends of the same pipeline. Supermemory is strongest at the front — getting heterogeneous data in, cheaply and de-duplicated, with connectors and profiles that assemble themselves. Genesys is strongest at the back — what happens to a memory after it is stored: how it links causally, how it ages, when it is forgotten, and whether you can explain any of that. This page compares the two honestly, and states the one benchmark number we can defend.
Memory-as-infrastructure, with an unusually wide ingestion surface.
Supermemory pitches a single API that swallows PDFs, web pages, images, and audio and turns them into a queryable memory stack with five layers: user profiles, a memory graph, hybrid retrieval, extractors, and connectors. The connector story is a real strength — Google Drive, Notion, OneDrive, Gmail, GitHub, and a web crawler feed memory without you writing ingestion code. Retrieval targets sub-300ms hybrid search, and the platform deduplicates at the token level, so re-uploading a document or re-syncing a connector does not redraw from your balance.
The open-source core is MIT-licensed and has real traction; the business is the hosted platform. Supermemory's memory graph tracks relationships with ontology-aware edges and handles knowledge updates, contradictions, and inferences automatically. If your problem is “ingest everything a user touches, across many content types and SaaS tools, and make it searchable with good user profiles on top,” Supermemory is built precisely for that shape of problem.
Ingestion breadth vs. causal depth.
The two products optimize different ends of the pipeline. Neither is universally better — the axis that matters is which pain you are trying to remove.
| Supermemory | Genesys | |
|---|---|---|
| Center of gravity | Ingestion + retrieval across many content types & connectors | Causal reasoning over one graph per user |
| Core unit | Memory + auto-built user profile / fact hierarchy | Memory node with typed causal edges |
| Graph | Ontology-aware edges (updates, contradictions, inferences) | Causal edges: caused_by, supersedes, context — walked by the scoring engine |
| Update model | Automatic contradiction / update handling in the graph | memory_amend supersedes; old node kept and decays, history queryable |
| Lifecycle | Profiles and hierarchies build in real time | Status FSM: TAGGED → ACTIVE → EPISODIC → DORMANT, with reactivation cascade |
| Forgetting | Token-level dedup; managed retention | Conjunctive prune: decay-score < 0.01 and orphaned and unpinned and not core |
| Explainability | Retrieved memories + profiles | memory_explain returns causal chain + live per-force score; memory_traverse returns edges |
Supermemory is strongest at the front: getting heterogeneous data in, cheaply and de-duplicated, with connectors and profiles that assemble themselves. Genesys is strongest at the back: how a memory links causally to what caused it, how it ages, when it is forgotten, and whether you can explain any of that. Supermemory's graph resolves contradictions automatically; Genesys resolves them legibly— a contradiction produces a supersede edge you can traverse, not a silent overwrite. If most of your engineering pain is ingestion breadth, Supermemory removes more of it. If your pain is “I cannot explain or audit what my agent remembers and why,” that is the Genesys axis.
One yardstick, no invented numbers.
Genesys is certified at 85.55 ± 0.37 on LoCoMo — the mean of ten independent runs under a frozen gpt-4o-mini answerer and judge at temperature 0 (n=1,540, categories 1–4; evaluated July 2026). The protocol is frozen and identical for every scored system: the answerer and judge are both gpt-4o-mini at temperature 0.
We could not source a Supermemory LoCoMo result under a comparable configuration, so this page does not put a number in Supermemory's mouth. We will not compare an unlike figure, and we will not invent one. For context, here are the two competitor numbers we can cite under the shared protocol, alongside ours.
| System | LoCoMo J (frozen protocol) |
|---|---|
| Mem0 (published, comparable) | 66.9 |
| Zep (corrected, comparable) | 75.14 |
| Genesys | 85.55 ± 0.37 |
The rule is symmetric. A LoCoMo number is comparable only when the answer model, the judge model, and the question set match. We exclude non-comparable self-reports on both sides — including our own modern-answerer run — rather than quote figures under different setups. The full protocol, the ten certified runs, and the failure ledger live on the LoCoMo benchmark page, with the complete method in the methodology dossier. If Supermemory publishes a LoCoMo number under this exact configuration, we will add it here verbatim.
Which to choose.
- Your dominant cost is ingestion breadth: PDFs, images, audio, web pages, and many SaaS connectors (Drive, Notion, Gmail, GitHub, S3) feeding one memory API.
- You want auto-built user profiles and fact hierarchies that assemble in real time without you modeling them.
- You care about token-level dedup because you re-sync the same sources repeatedly and do not want to pay twice.
- Your pricing shape is usage-based and you like the granularity — memory at $0.005/1K SM tokens (plain) or $0.010/1K (rich), search at $0.005/1K queries, with a $0 free tier that includes roughly $5/mo of usage.
- You need a fast hosted memory API and want SOC 2 / HIPAA and a self-hosted option on higher tiers (Scale $399/mo and Enterprise).
- Causal recall matters. You need to traverse why a memory exists and what it superseded — not just retrieve it.
memory_explainreturns the chain and a live score breakdown. - Memory lifecycle matters. Consolidation, dormancy, and reactivation cascades are first-class, so memory ages and revives instead of only accumulating.
- Explainability matters. Contradictions become inspectable supersede edges, not silent auto-resolutions; every prune obeys a conjunctive rule you can audit.
- You want a checkable number. Genesys is certified at 85.55 ± 0.37 under a frozen, published protocol with an open harness. Causal recall under a fixed judge is a benchmark we ran ten times.
- You want one portable, user-owned graph across ChatGPT, Claude, and anything speaking MCP — inspectable and theirs.
Where these facts come from.
Competitor facts were re-verified against primary sources — pricing page, homepage, docs, GitHub — on 2026-07-19. Pricing and usage rates change often; treat figures as of that date. Genesys numbers are the certified frozen-protocol figures at genesys.astrixlabs.ai/benchmarks/locomo.
Related.
LoCoMo results
The certified number and how to cite it, with the comparability caveats spelled out.
What is an AI memory layer?
Vector stores vs. temporal knowledge graphs vs. causal graphs — the honest tradeoffs.
The open-source engine
The causal-graph memory layer, AGPLv3, inspectable and portable over MCP.
Genesys overview
A memory you own and can inspect — connected once, read everywhere.