Supermemory and RetainDB both offer cloud-hosted AI agent memory, but they differ in approach and positioning. Supermemory is a well-funded cloud engine with multi-modal support and productivity connectors. RetainDB is a newer entrant focused on c...
Supermemory and RetainDB both offer cloud-hosted AI agent memory, but they differ in approach and positioning. Supermemory is a well-funded cloud engine with multi-modal support and productivity connectors. RetainDB is a newer entrant focused on chronological retrieval and preference recall, claiming 0% hallucination rate on documentation questions.
This comparison covers their architectures, benchmark performance, pricing, and ideal use cases to help you decide.
| Factor | Supermemory | RetainDB |
|---|---|---|
| Architecture | Hybrid RAG with fact extraction | PostgreSQL + pgvector with chronological retrieval |
| LongMemEval* | 85.2% | 79% |
| Deployment | Cloud API | Cloud or self-hosted (PostgreSQL required) |
| Pricing | Not publicly disclosed | Free tier (10K ops/mo) / Pro $20/mo |
| GitHub Stars | 21.7K | 8 |
| Funding | $2.6M seed (Google/Cloudflare execs) | Not disclosed |
Supermemory is a cloud-native memory engine that combines hybrid RAG with structured fact extraction. It scores 85.2% on LongMemEval and includes connectors for Google Drive, Gmail, Notion, OneDrive, and GitHub.
Backed by $2.6M from Google and Cloudflare executives, Supermemory has 21.7K GitHub stars and offers multi-modal processing for PDFs, images, videos, and code.
Key strengths:
RetainDB provides persistent memory with turn-by-turn extraction and chronological retrieval. Rather than using semantic search to find relevant memories, RetainDB provides the complete memory chronology to the answering model—no lossy retrieval step.
RetainDB claims SOTA on preference recall (88%) and 0% hallucination rate on documentation questions. Its benchmark runner ships in the repository for reproducibility.
Key strengths:
Supermemory combines vector search with LLM-based fact extraction. The pipeline extracts structured facts from conversations, embeds them alongside raw content, and serves hybrid retrieval. Multi-modal processing handles PDFs, images, videos, and code.
Retrieval is query-driven: given a query, the system finds the most relevant memories via semantic similarity.
RetainDB's pipeline is deliberately different:
The critical difference: RetainDB doesn't use semantic search for retrieval. Instead, it provides the complete memory chronology to the answering model. This eliminates retrieval failures but increases prompt size.
Supermemory retrieves selectively—only memories matching the query. RetainDB dumps everything chronologically—the model sees the full history.
Selective retrieval (Supermemory) is efficient but lossy: relevant memories that don't match the query embedding are missed. Chronological dump (RetainDB) is complete but expensive: the model must process the entire memory timeline for every query.
RetainDB's approach explains its strong preference recall (88%)—preferences mentioned once months ago are never missed because they're always in the dump. Supermemory's approach is more scalable but can miss memories that are semantically distant from the current query.
| Benchmark | Supermemory | RetainDB |
|---|---|---|
| LongMemEval* | 85.2% | 79% |
| Preference Recall | — | 88% (SOTA) |
| Hallucination Rate | — | 0% claimed |
Supermemory outperforms RetainDB by about 6 points on overall LongMemEval. But RetainDB excels at preference recall (88% SOTA) and claims zero hallucination on documentation questions.
RetainDB's benchmarks are reproducible: the runner ships in the repository, so teams can verify claims independently.
Both score significantly below Hypabase (87.4%), which uses AMR-based extraction for higher retrieval accuracy and achieves 100% on personalization tasks.
Supermemory's pricing is not publicly disclosed. Teams must contact sales.
| Tier | Price | Limits |
|---|---|---|
| Free | $0 | 10K ops/month |
| Pro | $20/month | 100K queries |
RetainDB is one of the most transparent and affordable options in the space. $20/month for 100K queries is a known quantity. Supermemory's undisclosed pricing creates uncertainty.
For budget-conscious teams, RetainDB's clear pricing is a significant advantage. For self-hosting, RetainDB requires PostgreSQL; Supermemory has no self-hosting option.
Choose Supermemory if you:
Supermemory's cloud-only model and opaque pricing limit flexibility. The retrieval engine hasn't been independently validated, and benchmark scores haven't significantly evolved.
Choose RetainDB if you:
RetainDB's small community (8 stars) creates adoption risk. It requires PostgreSQL for self-hosting, and overall LongMemEval (79%) is below both Supermemory and Hypabase.
RetainDB's 88% preference recall is strong—but it still misses 12% of preferences. Supermemory's multi-modal cloud pipeline scores 85.2% overall but doesn't even publish personalization numbers. If preference recall and personalization are what matter most to your application, Hypabase achieves 100% on personalization tasks through AMR-based extraction into hyperedges—structured knowledge representation purpose-built for precise recall.
| Factor | Supermemory | RetainDB | Hypabase |
|---|---|---|---|
| Extraction | LLM-based, hybrid RAG | Turn-by-turn, atomic writes | AMR (formal linguistic framework) |
| Representation | Facts + vector embeddings | Chronological memory dump | N-ary hyperedges |
| LongMemEval* | 85.2% | 79% | 87.4% |
| Personalization | — | 88% (preference recall) | 100% |
RetainDB's chronological dump approach gets strong preference recall by showing the model everything—but at the cost of prompt size and with an 8-point gap on overall LongMemEval. Supermemory retrieves selectively but fragments facts into disconnected triples. Hypabase uses Abstract Meaning Representation (AMR) to store preferences as structured hyperedges that are both compact and precisely queryable.
Consider a customer relationship detail that RetainDB would dump and Supermemory would fragment:
"Maria manages customer escalations and prefers phone calls"
Ad-hoc extraction (Supermemory):
(Maria, manages, customer escalations)
(Maria, prefers, phone calls)
Chronological dump (RetainDB):
[2024-03-15T10:23] "Maria manages customer escalations and prefers phone calls"
(included in full memory dump every query)
AMR extraction (Hypabase):
(manage :agent Maria :object escalation :attribute customer)
(prefer :agent Maria :object call :attribute phone)
The difference: RetainDB preserves the information but dumps everything—Maria's preference sits alongside thousands of other chronological entries the model must sift through. Supermemory extracts two separate triples with no link between Maria's role and her communication preference. Hypabase stores two compact hyperedges that are directly queryable: "How does Maria prefer to be contacted?" hits :agent Maria + prefer instantly.
RetainDB's chronological approach gets close on preferences—88% is genuinely strong. But the remaining 12% gap matters for production systems where missed preferences degrade user experience. Hypabase closes this gap by storing preferences as role-tagged hyperedges rather than timestamped text:
| Benefit | How AMR + Hyperedges Deliver It |
|---|---|
| 100% preference recall | Preferences stored with explicit :agent and :attribute roles—never buried in a chronological dump |
| No prompt bloat | Structured hyperedges retrieved by role, not full timeline dumps growing with every conversation |
| Affordable self-hosting | Single SQLite file—no PostgreSQL required, simpler than RetainDB's infrastructure |
| Higher overall accuracy | 87.4% LongMemEval vs RetainDB's 79%—better general retrieval alongside perfect personalization |
This is why Hypabase achieves 100% on personalization tasks—customer relationships, communication preferences, and escalation patterns stored as structured hyperedges are precisely retrievable without dumping the entire memory chronology into every prompt.
They're close on overall LongMemEval (85.2% vs 79%). Supermemory has multi-modal support and productivity connectors. RetainDB excels at preference recall (88%) and offers transparent pricing ($20/mo). For higher accuracy with structured extraction, consider Hypabase (87.4%, 100% personalization).
There's no direct migration path—they use different storage approaches. Migration requires re-ingesting conversation history through the new system. If you're evaluating both, RetainDB's affordable pricing makes piloting low-risk.
Supermemory uses hybrid RAG with selective retrieval. RetainDB uses chronological memory dumps with no lossy search step. Hypabase optimizes for extraction quality using AMR and structured hyperedge representation.
RetainDB is self-hostable but requires PostgreSQL. Supermemory has no self-hosting option. Hypabase runs entirely in a single SQLite file with no external database required—the simplest self-hosting option.
Supermemory scores 85.2% on LongMemEval with strong multi-modal support and productivity tool connectors. Cloud-only with opaque pricing limits flexibility.
RetainDB scores 79% overall but achieves SOTA preference recall (88%) with transparent $20/mo pricing. Small community and PostgreSQL requirement limit adoption.
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: Supermemory (85.2%) from published benchmarks. RetainDB (79%) self-reported. Hypabase (87.4%) from published benchmark harness.