Zep

Zep is a developer tool. Zep unifies business data, documents, and conversations into governed Context Graphs that AI agents query for task-specific facts. It combines temporal memory, sub-200ms retrieval, and policy-based access control, with cloud, BYOK, and VPC deployment options.

Zep is a context infrastructure layer built for teams running AI agents in production. It ingests business data, documents, and conversations, then compiles them into a Context Graph for each user, customer, account, or domain. Instead of pulling whole records through tool calls, an agent asks Zep for the facts a task depends on and receives a compact, policy-filtered response in a single call. The company says this cuts token use by around 90 percent compared with tool calling. The system is aimed at engineering and platform teams that need agents to act on accurate, current knowledge across an organization. Facts inside a graph are dated and updated as source data changes, and each fact retains a link back to where it came from, so an answer can be audited to its origin. A Context Lake holds many graphs at once, managed as one system, and retrieval stays under 200 milliseconds according to the vendor regardless of graph count or size. On the published LoCoMo and LongMemEval benchmarks, Zep reports 94.7 percent and 90.2 percent accuracy respectively with context sizes in the low thousands of tokens. Governance sits in the data layer rather than the application. Attribute-based access control determines which context an agent may retrieve, retention policies expire data on a schedule, legal holds block deletion, and request logs record every access decision. Deployments come in three forms: a managed cloud service, a cloud option where the customer supplies encryption keys, and a bring-your-own-cloud setup inside a customer VPC. Zep competes in the agent memory and retrieval space, where alternatives tend to be either vector stores with limited structure or tool-calling pipelines that pass entire records. Its differentiator is the temporal graph plus built-in policy enforcement. Pricing details are not spelled out on the homepage; enterprise plans and a demo request are the stated routes, with a self-serve start also offered.