ZepAgent memory platform for enterprise-scale AI, built on context graphs.

Zep is an enterprise-scale memory platform designed for AI agents, addressing the challenge of maintaining and managing agent memory across numerous users, business data, and past interactions. It aims to provide agents with a continuously learning and evolving understanding of their operational environment, thereby enhancing personalization and accuracy in agent interactions. The core of Zep's architecture is its proprietary Context Graph Engine, which constructs and manages a "Context Lake" of millions of individual context graphs. These graphs are built from diverse sources, including chat history, business data, and user interactions. Zep processes this information to create token-efficient, relevant context for agents through automated context assembly. A key capability is its temporal context graph, which automatically invalidates old facts when new information emerges, ensuring agents always reason with the most current data. Previous states are preserved as history, allowing agents to query what was true at any past date. This system also incorporates "Observations," where Zep analyzes graph structures to surface patterns, recurrences, and co-occurrences in memory, providing agents with a global perspective beyond isolated facts. Zep emphasizes enterprise-grade governance, offering features like attribute-based access control, policy-driven data retention, and full provenance tracking. Every fact within the graph traces back to its original source episode, enabling auditability. The platform is engineered for performance, demonstrating sub-200ms retrieval latency even with graph sizes up to 100 million entities. Designed for seamless integration, Zep can be added to existing agent frameworks or used independently, with SDKs available for Python, TypeScript, and Go. It aims to serve as a foundational layer in the enterprise agent stack, providing a scalable and governed solution for managing complex, evolving agent memory.