Last updated: March 2026 · Hypermemory is our product
Hypermemory vs Zep: AI Memory Comparison
Zep (powered by its open-source Graphiti engine, 24K+ stars) introduced temporal knowledge graphs to AI memory — a genuine architectural innovation. Hypermemory extends this with six retrieval strategies, predictable pricing, and instant ingestion. Here's the full comparison.
Choose Hypermemory if…
- →You need instant post-ingestion retrieval (not hours later)
- →Predictable pricing matters — no credit consumption surprises
- →You want to self-host without provisioning Neo4j/FalkorDB
- →Benchmark accuracy (LoCoMo SOTA) is a requirement
Choose Zep if…
- →Your primary use case is CRM agents with deep relationship graphs
- →You need BYOC (bring your own cloud) into your AWS VPC
- →You prefer Graphiti's raw open-source engine for custom builds
Feature-by-Feature Comparison
Pricing data as of March 2026. Sources: getzep.com/pricing, hypermemory.run/pricing.
| Feature | Hypermemory | Zep |
|---|---|---|
| Semantic Search | ✅ | ✅ |
| BM25 / Keyword Search | ✅ | ✅ |
| Temporal Scoring (recency ranking) | ✅ | ✅ |
| Temporal Fact Search (date-aware queries) | ✅ | ✅ |
| Fact Matching (entity-attribute-value triples) | ✅ | ✅ |
| Multi-hop Reasoning | ✅ All tiers | ✅ |
| Temporal Supersession (contradictions) | ✅ | ✅ |
| Lazy Ingestion (instant add, async extraction) | ✅ | ❌ Delayed (hours) |
| Retrieval Strategies | 6 strategies fused | 3 strategies (semantic + BM25 + graph) |
| LoCoMo Benchmark | 87–94% (SOTA) | ~85% |
| Free Tier | ✅ 10K memories | 1,000 credits only |
| Entry Paid Plan | $19/mo (all features) | $25/mo Flex |
| Mid-tier Pricing | $19/mo (all features) | $475/mo (300K credits) |
| Predictable Pricing | ✅ Fixed/tier | ❌ Credit consumption varies |
| Self-Hosting | ✅ Single system | ⚠️ Partial Graphiti + graph DB + embeddings |
| Self-Host Complexity | Low | High (3+ systems) |
| Community Edition | ✅ | ❌ Deprecated |
| TypeScript SDK | ✅ | ✅ |
| Framework-Agnostic | ✅ | ✅ |
| MCP Integration | ✅ | ⚠️ Partial |
| SOC 2 / HIPAA | ✅ Enterprise | ✅ Enterprise |
| On-Premise (BYOC) | ✅ | ✅ Enterprise |
The Delayed Retrieval Problem
Zep's core architectural trade-off is background graph processing. When you ingest new memories, Zep extracts entities, resolves relationships, and rebuilds temporal graph edges asynchronously. This is powerful for long-term accuracy — but it means memories ingested moments ago may not be retrievable for hours.
Hypermemory's lazy fact extraction model solves this: raw content is stored and immediately searchable. Entity extraction and temporal fact building happen in the background without blocking retrieval. Your agent always has access to the latest context, even before background processing completes.
Pricing: Predictable vs Credit-Based
Zep's credit model charges per "episode" (a conversation turn or document). Episodes over 350 bytes are billed in multiples. High-volume agentic workloads — where agents process many turns per session — can spike credit consumption unpredictably.
$19/mo
Hypermemory Pro
Fixed. All 6 strategies.
$25/mo
Zep Flex
20,000 credits. Vector + graph.
$475/mo
Zep Flex Plus
300,000 credits + overages.
Zep's free tier offers only 1,000 credits — insufficient for meaningful prototyping. Hypermemory's free tier includes 10,000 memories with no expiration.
Self-Hosting Complexity
Zep's community edition — the self-hosted option — was deprecated. Self-hosting Zep now requires running three separate systems:
- The raw Graphiti library (no managed installer)
- A graph database: Neo4j 5.26, FalkorDB 1.1.2, Kuzu 0.11.2, or Amazon Neptune
- Your own embedding infrastructure
Hypermemory ships as a single deployable system. On-premise deployment is available on the Enterprise plan with dedicated infrastructure support.
Try Hypermemory Free
10,000 memories, all six retrieval strategies, instant ingestion. No credit card required.