Zep is a temporal knowledge graph platform that gives enterprise AI agents long-term memory and business context. Built by Zep AI, Inc., it ingests documents, conversations, JSON records, and application events.
Standard RAG relies on static vector search over text chunks. Zep instead builds a Context Graph per user, customer, account, or domain, and tracks how facts change over time.
Incoming data arrives as “Episodes” and is processed by Graphiti, Zep’s open-source framework. Graphiti extracts entities, relationships, and temporal metadata. Konig, Zep’s graph database service, serves queries in under 200 milliseconds.
Access control, retention policies, and data provenance are enforced in the data layer. This makes Zep a fit for teams that need auditable agent memory.
Pricing
The free tier includes 10,000 credits per month and 2 projects, but credits do not roll over. Flex costs $125 per month for 50,000 credits, and Flex Plus costs $375 per month for 200,000 credits, with 17% savings on annual billing. One credit covers each Episode up to 350 bytes, plus one credit per additional 350 bytes, while retrieval and storage cost 0 credits. Enterprise pricing is custom and negotiated with sales.
* Disclaimer: Please note that pricing information may not be up to date. For the most accurate and current pricing details, refer to the official website.
Key Features
- ✓
Bi-temporal edges record when facts were valid
- ✓
Hybrid retrieval merges vector, BM25, and graph search
- ✓
Attribute-based access control on reads and writes
- ✓
Retention policies with legal hold support
- ✓
Isolated graphs per user, customer, or account
Use Cases
Customer Account Support Agents
Zep merges CRM records, Slack messages, and support logs into one customer graph. A support agent can retrieve past escalations and contract status in under 200 milliseconds. This reduces conflicting answers during ticket resolution.
Long-Term Assistant Memory
A chat assistant stores user preferences and past interactions in a dedicated user graph. Over days or weeks, Zep updates that graph and injects only the needed context into prompts. Users get continuity without oversized prompts.
Shared Memory Across Agents
Procurement, renewals, and billing agents read and write verified facts in a shared Context Lake. Access rules limit what each agent can view or change. Teams share knowledge without exposing restricted data.
Audit Trails in Regulated Industries
Financial services and healthcare teams must trace AI outputs to their sources. Compliance staff follow an agent answer back to the original document or message. They can also apply retention schedules and legal holds.
Persistent Memory for Coding Assistants
Developers connect Claude or Cursor to Zep’s Context MCP Server. The coding assistant gains graph-backed memory without changes to the existing agent runtime or orchestration layer.
Strengths & Weaknesses
Strengths
The bi-temporal model keeps history for every fact as information changes.
Retrieval runs in one sub-200ms query without LLM reranking overhead.
Konig keeps active graphs in RAM and moves inactive ones to cheaper storage.
Provenance tracing links each extracted fact to its source message or document.
Deployment options include managed cloud, customer-managed keys, and private VPC.
Weaknesses
Self-serve plans cap projects, credits, custom types, and API log retention at 1 to 7 days.
Webhooks, analytics, and custom extraction instructions require Flex Plus or higher.
HIPAA BAAs and one-year audit logs are limited to Enterprise plans.
Graph and temporal concepts create a learning curve for teams used to basic RAG.
Who Is This For?
Enterprise AI engineers and architects: They need production memory infrastructure for multi-agent systems without building a custom graph database.
SaaS and AI application developers: They want SDKs and MCP support to add persistent user memory to conversational assistants.
Security and compliance officers: They need access policies, SOC 2 Type II reporting, and audit logging around agent data.
Customer success and support operations teams: They benefit from agents that combine past conversations, CRM records, and agreements.
Frequently Asked Questions
What counts as an Episode in Zep?
An Episode is any single data object sent to Zep. Examples include a chat message, a JSON payload, or a block of text.
What is Graphiti and how does it relate to Zep?
Graphiti is an open-source context graph framework that originated at Zep. It extracts entities, relationships, and temporal metadata from Episodes.
Which compliance certifications does Zep hold?
Zep is SOC 2 Type II certified. HIPAA Business Associate Agreements are available only on the Enterprise plan.
Can EU companies sign a data processing agreement?
Yes. Zep signs Data Processing Agreements with EU customers and customers with an EU presence.
How accurate is Zep on memory benchmarks?
Zep reports 94.7% accuracy on LoCoMo at 155 ms latency. On LongMemEval it reports 90.2% accuracy at 162 ms.
How does Zep differ from a standard RAG setup?
Standard RAG searches static text chunks. Zep tracks how facts change over time and combines vector, BM25, and graph retrieval.
How hard is it to get started?
You create a project, add the SDK, and send data as Episodes. Teams new to graph concepts should expect a learning curve.
Which deployment options are available?
Options include managed Zep Cloud, Zep Cloud with customer-managed encryption keys, and Bring Your Own Cloud inside a private VPC.
How long are API logs retained?
Self-serve plans keep API logs for 1 to 7 days. Enterprise plans retain detailed audit and API logs for up to 1 year.
Zep works with agent frameworks including LangGraph, Vercel AI SDK, Pydantic AI, and Google ADK. Its Context MCP Server connects to Claude and Cursor. Graphiti supports Neo4j, FalkorDB, and Amazon Neptune as graph backends. Supported LLM backends are OpenAI, Azure OpenAI, Google Gemini, and Anthropic.