Zep (agent memory)
Zep is a memory layer for AI agents that builds a temporal knowledge graph from conversations and business data; its open-source core is called Graphiti and is licensed Apache 2.0.
Zep gives agents long-term memory in the form of a temporally versioned knowledge graph. Facts are stored as relationships between entities with a validity period — outdated statements are not deleted, they become invalid.
Explanation
A chat history in the context window is not memory, it is a note that eventually gets truncated. Zep extracts entities and facts from conversations and connected systems and stores them as nodes and edges — "customer uses plan B" instead of 40,000 characters of raw transcript. Every edge carries timestamps for when validity starts and ends. So the agent does not ask "what was in the transcript" but "what holds true about this person today".
Marketing Relevance
In marketing and service, memory is the difference between a chat and a relationship. An agent that knows what was promised last time saves follow-up questions. That is exactly why this layer is sensitive under data protection law: it is deliberately a personal profile and needs purpose limitation, a retention period and a deletion path.
Example
A prospect asks about pricing in the website chat and about availability a month later. The agent recognises the same person, knows an offer is open and, instead of a canned answer, suggests a call with the responsible consultant.
Common Pitfalls
A memory that keeps everything is a liability: without a retention period, a profile nobody asked for keeps growing. Second point: extracted facts are model output and therefore fallible — treating them as truth without review preserves misunderstandings for months.
Origin & History
Zep started as a memory service for LLM applications; in 2024 Graphiti followed as an open-source framework for temporal knowledge graphs, which now forms the basis of the service. Agent memory moved from side question to architectural decision in 2025/2026.
Further Resources
Marketing Use Cases
Recognition and context across multiple website chat conversations.
Service cases where commitments from earlier contacts still apply.
Personalisation based on verified facts instead of raw transcripts.
Frequently Asked Questions
Why not just put the whole transcript in the prompt?
Because it is expensive, slow and eventually impossible. A graph delivers the few relevant facts instead of hundreds of messages — and it can say what no longer holds.
Is Graphiti the same as Zep?
No. Graphiti is the open-source framework for the temporal knowledge graph, Zep the service built on top. If you need data sovereignty, you can self-host Graphiti.
Related Services
Go deeper: Agentic AI Hub · Model comparison 2026