Zep
Context engineering platform for AI agents with unified context graph, Graph RAG, and automated context assembly. 200ms retrieval.
Last verified:
What is Zep?
Zep is an end-to-end context engineering platform that delivers the right information at the right time with sub-200ms latency for AI agents. It serves as a memory foundation for your AI stack, providing long-term memory storage and enrichment for AI assistant apps. Zep solves the agent context problem by building a temporal knowledge graph that learns from user interactions and business data, enabling AI assistants to provide personalized, accurate, and up-to-date information.
Key features include the Context Graph Engine that constructs graphs from any source (chat history, business data, user interactions), automated context assembly, and the Context Lake that manages millions of context graphs as one governed system. Zep provides sub-200ms retrieval latency regardless of graph size, attribute-based access control, policy-driven retention with legal hold capabilities, and comprehensive audit logs. The platform also includes built-in observability for latency, error rate, and retrieval activity, plus analytics for usage and reliability tracking.
Zep is designed for enterprise AI teams building reliable, personalized agents. It works with any agent framework or none, requiring just three lines of code to add memory. The platform is ideal for startups and enterprises that need agents to recall past conversations, reduce hallucinations, and reason about how facts change over time. Teams at Torq, AlphaSignal, Flockx, and Axtria use Zep in production for real-world agent use cases.
Zep pricing
Pricing model: Free
Zep offers a free tier with 1,000 credits/month, low rate limits, and lower priority episode processing. The Flex Plan is $25/month for 20,000 credits, 600 requests per minute, 5 projects, 10 custom entity and edge types, and unlimited memories, retrieval, and users. Additional credits available at $25 per 20,000. Flex Plus is $475/month with 300,000 credits, 1,000 requests per minute, 5 projects, 20 custom entity and edge types, custom extraction instructions, webhooks, and 7-day API logs. Additional credits at $125 per 100,000. Enterprise pricing is not publicly listed and includes advanced security, compliance, and flexible deployment options. Unused credits roll over for up to 60 days. HIPAA BAA available as add-on.
Zep pros
- Sub-200ms retrieval latency regardless of graph size
- Temporal knowledge graph tracks how facts change over time
- Works with any agent framework or none
- Only three lines of code to add memory
- Combines chat history with structured business data
- Automatic fact invalidation when new information contradicts old
- Attribute-based access control for context access
- Policy-driven retention with legal hold capabilities
- Complete data provenance tracing to source episodes
- Observability built in for latency and error rates
- SOC 2 Type II and HIPAA BAA compliance
- Three deployment models: Cloud, BYOK, and BYOC
- Surfaces Observations: patterns and co-occurrences in memory
- Token-efficient context assembly reduces prompt costs
- Group-level graphs capture organizational knowledge
- Python, TypeScript, and Go SDKs available
- Open-source Community Edition available
Zep cons
- Free tier limited to 1,000 credits per month
- Credit-based pricing may be unclear for new users
- Data size limit of 10,000 characters per graph.add call
- Maximum 10 metadata keys per episode
- Metadata only supports scalar types (no nested objects)
- Does not process text directly from links or files
- Enterprise pricing not publicly listed
- Lower priority episode processing on free tier
Frequently asked questions about Zep
What is Zep and what does it do?
Zep is a context engineering platform that delivers agent memory at enterprise scale. It builds a temporal knowledge graph (Context Graph) from any source—chat, business data, documents, JSON—and serves token-efficient, prompt-ready context from a governed Context Lake with sub-200ms retrieval. This enables reliable, personalized agents with fewer hallucinations and higher accuracy.
Is there a free version of Zep Cloud?
Yes, Zep offers a free tier with 1,000 credits per month, low rate limits, and lower priority episode processing.
How does Zep handle facts that change over time?
When new information contradicts what's in the graph, Zep invalidates the old fact. Your agent reasons with the latest decisions, traits, and behaviors. Old facts stay as history, so you can ask what's true now or what was true on any past date. Each fact includes valid_at and invalid_at dates.
What deployment models does Zep support?
Zep offers three deployment models: Cloud (managed service with no infrastructure to run), BYOK (Cloud with your own encryption keys where you control the keys), and BYOC (Bring Your Own Cloud, deployed inside your VPC with your network and compliance boundary).
WhatSDKs does Zep provide?
Zep offers Python, TypeScript, and Go SDKs. You can install via pip install zep-python for Python, npm install @getzep/zep-js or yarn add @getzep/zep-js for TypeScript.
What is the Context Lake?
The Context Lake is Zep's proprietary system that manages millions of context graphs, governed and served as one system. It applies the data-lake pattern to agent context, powered by Zep's Context Graph Engine. It provides sub-200ms retrieval regardless of graph size or count, with authorization, retention, and audit living in the substrate.
What are Observations in Zep?
Observations are patterns, recurrences, and co-occurrences in memory that Zep surfaces by analyzing the structure of the graph. Your agent gains a global perspective beyond facts and summaries, enabling it to understand patterns like preferences or behavioral correlations.
How do I add business data to Zep?
Zep supports three data types: message (chat messages not associated with a Thread's chat history, like emails), text (raw text without speaker attribution, like documents or wiki articles), and JSON (REST API responses or JSON-formatted business data). You add data via the graph.add method with a graph_id or user_id, and can attach up to 10 key-value metadata tags.
What compliance certifications does Zep have?
Zep Cloud is SOC 2 Type II certified and offers HIPAA BAA (Business Associate Agreement) as an add-on. The BYOC deployment model lets you maintain your own compliance boundary.
Does Zep work with LangChain?
Yes, Zep has LangChain support including a ZepChatHistory class for native persistence and retrieval of chain and agent memory, plus a ZepRetriever class enabling contextual vector search over a user's long-term memory to provide agents with historical context.