Zep and Supermemory are both cloud-focused memory solutions, but they take different architectural approaches. Zep uses a temporal knowledge graph (Graphiti) with time as a first-class dimension, while Supermemory combines hybrid RAG with structur...
Zep and Supermemory are both cloud-focused memory solutions, but they take different architectural approaches. Zep uses a temporal knowledge graph (Graphiti) with time as a first-class dimension, while Supermemory combines hybrid RAG with structured fact extraction and multi-modal support.
This comparison covers their architectures, benchmark performance, pricing, and ideal use cases to help you decide.
| Factor | Zep | Supermemory |
|---|---|---|
| Architecture | Temporal knowledge graph (Graphiti) | Hybrid RAG with fact extraction |
| LongMemEval* | 71.2% | 85.2% |
| Deployment | Cloud-first; self-host requires Graphiti + graph DB | Cloud API only |
| Pricing | Free / $25 / $475 / Enterprise | Not publicly disclosed |
| GitHub Stars | 4.4K (Zep) + 24.8K (Graphiti) | 21.7K |
| Funding | Not disclosed | $2.6M seed (Google/Cloudflare execs) |
Zep uses Graphiti, a temporal knowledge graph where time is a first-class dimension. Every fact has valid_from, valid_to, and invalid_at markers, allowing queries like "what was true in January?" or "when did this change?"
Zep positions itself around "context engineering" rather than just memory. Graphiti, the underlying engine, has 24.8K stars and supports multiple graph backends (Neo4j, FalkorDB, Kuzu, Neptune).
Key strengths:
Supermemory uses a hybrid RAG approach that combines vector search with LLM-based fact extraction. Founded by a 19-year-old and backed by $2.6M from Google and Cloudflare executives, it focuses on fast retrieval (~50ms) and multi-modal support.
Beyond basic memory, Supermemory includes connectors for Google Drive, Gmail, Notion, OneDrive, and GitHub—positioning it as a broader knowledge integration layer.
Key strengths:
Zep's Graphiti engine stores facts as nodes in a knowledge graph with explicit temporal metadata. Each edge carries validity windows tracking when facts became true and when they were superseded.
This temporal awareness is native to the architecture—not a filter applied after retrieval. The tradeoff is infrastructure complexity: self-hosting requires running Graphiti plus a graph database.
Supermemory combines vector search with LLM-based fact extraction. The memory engine extracts facts, handles temporal changes and contradictions, and builds user profiles with ~50ms retrieval. It also processes multi-modal content: PDFs, images (OCR), videos (transcription), and code.
The hybrid approach blends RAG-style retrieval with structured memory queries, aiming for both speed and accuracy.
Zep structures memory around temporal relationships; Supermemory structures memory around multi-modal integration.
When a user changes jobs, Zep's temporal edges explicitly mark the old job as superseded with invalid_at timestamps. The graph knows exactly when the change happened and can query historical state.
Supermemory's fact extraction handles the same update through contradiction resolution—detecting that the new employment fact conflicts with the old one. This is effective but less architecturally explicit than Zep's temporal model.
The tradeoff: Zep is more precise for time-based queries, while Supermemory handles a wider range of input types and retrieves faster.
| Benchmark | Zep | Supermemory |
|---|---|---|
| LongMemEval* | 71.2% | 85.2% |
Supermemory outperforms Zep by 14 percentage points on LongMemEval. Zep's 71.2% is self-reported; independent evaluation shows 63.8%, which would widen the gap to over 21 points.
Both score significantly below Hypabase (87.4%), which uses AMR-based extraction for higher retrieval accuracy.
| Tier | Price | Limits |
|---|---|---|
| Free | $0 | 1K episodes/month |
| Flex | $25/month | 20K credits, 600 req/min |
| Flex Plus | $475/month | 300K credits, 1K req/min, webhooks |
| Enterprise | Custom | SOC2, HIPAA, dedicated support |
Supermemory's pricing is not publicly disclosed. Contact their team for details.
Zep's transparent pricing is an advantage for budgeting. Supermemory's opaque pricing makes cost comparison difficult, though its cloud-only model means you avoid infrastructure costs for graph databases that Zep self-hosting requires.
Choose Zep if you:
Zep's temporal graph is useful for knowledge updates, though the architecture hasn't evolved much since launch.
Choose Supermemory if you:
Supermemory scores higher on benchmarks but is cloud-only with undisclosed pricing, making it harder to evaluate total cost of ownership.
Zep and Supermemory both rely on LLM-prompted extraction—whether through graph triples or hybrid RAG—which means the quality of stored facts depends entirely on how well the LLM prompt generalizes. When facts involve multiple attributes about a person (preferences, location, communication style), ad-hoc extraction routinely scatters them across disconnected records. Hypabase solves this with formal linguistic extraction.
| Factor | Zep | Supermemory | Hypabase |
|---|---|---|---|
| Extraction | LLM-based into triples | LLM-based hybrid RAG | AMR (formal linguistic framework) |
| Representation | Temporal triples | Vector + extracted facts | N-ary hyperedges |
| LongMemEval* | 71.2% | 85.2% | 87.4% |
| Personalization | — | — | 100% |
Hypabase uses Abstract Meaning Representation (AMR)—a formal framework from computational linguistics—to map sentences into structured graphs. The output uses PENMAN notation with karaka semantic roles, capturing every relationship in a single atomic fact:
"Bob prefers Slack over email and works remotely from Denver"
Ad-hoc extraction (Zep, Supermemory):
(Bob, prefers, Slack)
(Bob, dislikes, email)
(Bob, works, remotely)
(Bob, location, Denver)
— Four separate facts; "Slack *over* email" preference ranking lost
AMR extraction (Hypabase):
(prefers :subject Bob :object Slack :origin email :attribute communication)
(works :subject Bob :attribute remote :locus Denver)
Hypabase preserves the preference ranking (Slack over email) and the compound attribute (remote work from Denver) as structured hyperedges. Ad-hoc extraction flattens "prefers Slack over email" into two disconnected facts, losing the comparison entirely.
| Benefit | How AMR + Hyperedges Deliver It |
|---|---|
| Preference rankings preserved | :object and :origin roles capture "X over Y" comparisons atomically |
| Multi-attribute facts | Location + work style stored together, not as separate unlinked records |
| Precise retrieval | Query "How does Bob prefer to communicate?" hits :subject Bob + :attribute communication |
| Consistent extraction | 6 karaka roles cover all semantic relationships—no ad-hoc relation types |
This is why Hypabase achieves 100% on personalization tasks—user preferences with nuanced rankings and compound attributes are precisely retrievable, even when queries use different phrasing.
Supermemory scores higher on LongMemEval (85.2% vs 71.2% self-reported). Zep offers stronger temporal reasoning and enterprise compliance. For higher accuracy with structured extraction, consider Hypabase (87.4%).
There's no direct migration path—they store different data structures. Migration requires re-ingesting your conversation history through the new system. If you're evaluating both, consider running a small pilot before committing.
Zep optimizes for temporal accuracy with graph-based reasoning. Supermemory optimizes for multi-modal integration and retrieval speed. Hypabase optimizes for extraction quality using AMR and structured hyperedge representation.
Neither is ideal. Zep requires running Graphiti plus a graph database (Neo4j/FalkorDB/Kuzu). Supermemory is cloud-only with no self-hosting option. Hypabase runs entirely in a single SQLite file with no external database required—the simplest self-hosting option.
Zep adds temporal reasoning but self-reports 71.2% on LongMemEval (independent evaluation shows 63.8%). Useful for knowledge updates, though requires graph database infrastructure.
Supermemory scores 85.2% with multi-modal support and fast retrieval, but is cloud-only with undisclosed pricing—harder to evaluate for budget-conscious teams.
Hypabase achieves 87.4% through AMR-based extraction into hyperedges—structured knowledge representation that preserves relationships ad-hoc extraction fragments. 100% on personalization tasks.
All three are straightforward to integrate:
LongMemEval scores: Zep (71.2%) self-reported; independent evaluation shows 63.8% (arxiv:2512.13564). Supermemory (85.2%) from published benchmarks. Hypabase (87.4%) from published benchmark harness.