MemoryAtlas

Context Engineering Platform (Zep v3) built on Graphiti — a bi-temporal knowledge graph that models facts as triplets (entity-relationship-entity) with temporal metadata on every edge. Contradictions are resolved via temporal edge invalidation, not LLM judgment: when a fact changes, the old edge is superseded and a new one is created, preserving the full historical timeline. Ingestion uses discrete episodes (text or structured JSON) that maintain data provenance and enable incremental extraction. Zep Cloud delivers enterprise-grade managed hosting (SOC 2 Type II, HIPAA BAA, BYOC); Graphiti is the Apache-2.0 OSS engine underneath.

Deep diveZep deep dive: how Graphiti's bi-temporal knowledge graph actually worksRead →
Storage
Zep Cloud (fully managed, three modes: Cloud / Cloud+BYOK / BYOC); OSS: Neo4j 5.26+, FalkorDB 1.1.2+ (Redis-based), Amazon Neptune (Database Cluster or Analytics Graph); Kuzu 0.11.2 supported but deprecated (upstream unmaintained as of v0.29.x)
Retrieval
Hybrid: vector similarity (semantic) + BM25 full-text search + graph traversal, with distance-based reranking from central entity nodes — no LLM-in-loop reranking step. Supports point-in-time historical queries via archived superseded edges. Zep Cloud p95 < 200 ms at up to 100M graphs.
Self-host
Self-host: heavy
License
Apache-2.0
Pricing
Zep Cloud: free tier (10,000 credits/month, 2 projects); Flex $1,250/yr (50K credits/month, $25/10K overage); Flex Plus $3,750/yr (200K credits/month, webhooks + custom extraction); Enterprise custom (SOC 2 Type II, HIPAA BAA, BYOC). Credit cost: 1 credit per episode up to 350 bytes; retrieval and storage free. Graphiti OSS engine free. · Free + paid
GitHub stars
29,735
Last release
2026-07-27
Last commit
2026-08-05
First catalogued
2026-06-28

Strengths

  • Bi-temporal edge invalidation resolves contradictions without LLM judgment — superseded facts are archived not deleted, enabling exact point-in-time historical queries
  • Hybrid retrieval (vector + BM25 + graph traversal) with distance-based reranking from central nodes avoids LLM-in-loop latency overhead
  • Broad graph DB support: Neo4j, FalkorDB (Redis-based), Amazon Neptune — choose backend by operational preference and cloud environment
  • Zep Cloud hits sub-200ms p95 retrieval at 100M graph scale and reports 94.7% on LoCoMo; SOC 2 Type II + HIPAA BAA for regulated workloads
  • Episodic ingestion with per-episode provenance; accepts both unstructured text and structured JSON in the same graph

Watch out

  • Self-hosting requires a dedicated graph DB (Neo4j 5.26+ or FalkorDB + Redis) plus LLM and embedding provider config — significant operational footprint; Zep Community Edition was deprecated
  • Kuzu backend officially deprecated in v0.29.x (upstream unmaintained); only Neo4j, FalkorDB, and Amazon Neptune are actively supported
  • Zep Cloud credit cost is episode-driven (1 credit per 350 bytes ingested); high-ingestion agents can exhaust the free 10K credits/month quickly — cost forecasting requires estimating episode volume

Best for

  • Agents tracking facts that change over time — customer profiles, evolving relationships, product details that must stay current without losing history
  • Enterprise multi-agent deployments needing SOC 2 / HIPAA compliance, BYOC deployment, and sub-200ms retrieval SLAs
  • Workflows requiring point-in-time historical queries ("what did the agent know about X on date Y?")

How it integrates

Benchmark results

BenchmarkValueBackboneTrustSource
longmemeval63.8 accuracygpt-4o-miniSelf-reportedZep AI (Rasmussen et al.)
longmemeval90.2 accuracygpt-5.4 (reasoning=medium)Self-reportedZep
longmemeval71.2 accuracyGPT-4oIndependentHindsight/Vectorize (competitor re-run)
longmemeval58.6 accuracygpt-4o-miniIndependentLiCoMemory (Huang et al., HKUST et al.) — competitor re-run
locomo44.76 accuracygpt-4o-miniIndependentLiCoMemory (Huang et al., HKUST et al.) — competitor re-run
locomo94.7 accuracygpt-5.4 (reasoning=medium)Self-reportedZep

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards