An agent memory and context-engineering platform built on a temporal knowledge graph. When information changes or conflicts, Zep invalidates outdated facts while preserving historical context, which suits multi-session agent workflows better than pure vector retrieval. It exposes retrieval and storage via API plus a hosted MCP Server, and reports 94.7% retrieval accuracy on LoCoMo (155ms) and 90.2% on LongMemEval (162ms) using single-pass retrieval rather than slow agentic tool loops. Pricing is credit-based at 1 credit per 350 bytes of episode data; retrieval, storage, and user counts are not billed separately. Free tier includes 10,000 credits/month; paid Flex starts at $125/month (50,000 credits, ~$104/month annual), Flex Plus at $375/month (200,000 credits), and Enterprise adds SOC 2 Type II, HIPAA BAA, BYOK, and VPC deployment. The core framework Graphiti is open source, while the hosted service is commercial. Early-stage company (YC W24), with customers reported to include Samsung, Zscaler, and HoneyBook.
**Pros & Cons**:
Pros:
Temporal knowledge graph handles fact invalidation and history retention, a real differentiator over plain vector stores; single-pass retrieval is fast (155–162ms) and avoids multi-step agentic tool loops; hosted MCP Server reduces integration work for agent frameworks; Graphiti is open source, so teams can self-host or inspect the core; SOC 2 Type II and HIPAA BAA available on Enterprise; pricing bundles storage/retrieval/users rather than charging per query.
Cons: Steep jump from the free 10,000 credits to the $125/month Flex tier with no mid-tier option; byte-based episode billing plus automatic top-up (10,000 credits for $25 when balance drops below 20%) means monthly spend has no hard ceiling; self-hosting Graphiti requires Docker, FalkorDB/Neo4j, and LLM key management; hosted MCP Server is billed per seat and custom OIDC/ABAC requires Enterprise; company is very early-stage (YC W24, ~$2.3M–$3.5M raised, small team), so enterprise delivery is still being built out.
**Who is it for?**:Developers building multi-session AI agents that need persistent, evolving user memory; teams that want fact-invalidation logic without building their own knowledge graph; agent frameworks looking for an MCP-native memory backend; enterprises requiring SOC 2 or HIPAA with BYOK/VPC options.