Agent Context
Is Hard.
Zep Fixed It.

The Problem No One Wanted to Admit

Your AI Agent Has the Memory of a Goldfish

Every enterprise demo looks identical. A sleek agent interface, a confident pitch, a few impressive responses - and then a customer asks a follow-up from last week's conversation and the whole thing falls apart. The agent doesn't remember. It never did. It was pretending.

This is the dirty secret of the AI agent boom. Models got dramatically better at reasoning, but nobody solved the basic problem of context. How does an agent know what happened before? How does it know what changed? How does it handle the fact that what was true in January might not be true in April?

Vector databases help. Retrieval-augmented generation helps. Shoving everything into a long context window - well, that works until you're paying $28 per conversation in tokens and your latency looks like a dial-up modem.

Daniel Chalef had seen this movie before. As the founder of KnowledgeTree in the early 2000s - one of the original open-source knowledge management systems - he understood what happens when information architecture gets messy at scale. When he came back to the problem of agent memory, he didn't reach for the nearest vector database. He reached for a graph.

Origin Story

Three Engineers, One Obsession, Zero Patience for Half-Solutions

Chalef founded Zep AI in 2023 alongside Preston Rasmussen and Pavlo Paliychuk - engineers he'd worked alongside in the trenches at companies like SparkPost, where he led ML engineering before the company was acquired by MessageBird. These weren't people who needed to discover the problem. They'd lived inside it.

The team joined Y Combinator's Winter 2024 batch with a clear thesis: the memory layer for AI agents is broken, and we know exactly how to fix it. They weren't building another chatbot wrapper. They were building the plumbing - the infrastructure that would make every other AI agent smarter by default.

The results came fast. By June 2024 - roughly three months into their YC batch, with a team of five - Zep hit $1M ARR. Not a sign-of-intent ARR. Not a letter of intent. Real revenue from paying customers. In the W24 cohort, that number was remarkable enough to turn heads in the YC network.

The investors who caught on early include Y Combinator (the $500K pre-seed), Root Ventures, Engineering Capital, and Step Function Ventures, plus angels from Vercel, Google, and Airtable - a signal list that reads like a who's who of people who understand developer infrastructure.

The Technology

A Graph That Remembers When Things Were True

Most memory systems for AI agents are essentially fancy search boxes. You chunk text, embed it, store it in a vector database, and hope you retrieve the right piece when the agent needs it. It works well enough for static facts. It falls apart the moment your world changes.

Zep's approach is fundamentally different. Their open-source engine, Graphiti, is a temporal knowledge graph - it doesn't just track what is true, it tracks when something was true and when that changed. When a customer switches pricing plans, when a patient's medication changes, when a user's preferences evolve - the graph knows, and the old information doesn't quietly persist to confuse future answers.

Why Graphs Beat Vectors for Agent Memory

Relationships matter in memory. "Alice is the VP of Engineering at Acme Corp, and she's been skeptical about the API integration since Q3" is not a chunk of text to retrieve - it's a web of entities and their changing relationships over time. A graph captures that structure natively.

Graphiti ingests chat history, business data, and external events, then assembles context blocks that are pre-formatted for LLMs - ready to drop straight into a prompt. The whole retrieval process happens in under 200ms at P95, which matters enormously for voice AI applications where every millisecond of hesitation sounds like the robot is buffering.

In the Wild

From Healthcare to E-Commerce: Who's Running on Zep

The customer list is a Rorschach test for where AI agents are actually getting deployed. Healthcare companies like Twin Health and Thrive AI Health use Zep for patient context and longitudinal tracking - where the cost of a forgotten medication is not a UX problem, it's a liability. Praktika.ai uses it for adaptive language learning. Swiggy, the Indian food delivery giant, runs it for personalized shopping agents. Torq uses it for security automation workflows.

Over 240 customers total, including Fortune 500 companies who don't make the public list. The 50% month-over-month ARR growth suggests the pipeline isn't slowing down.

The Open Source Bet

25,000 Stars and a Research Paper to Back It Up

Most enterprise AI startups treat open source like a marketing checkbox. Zep built it like they meant it. Graphiti - the temporal graph engine at the heart of everything - is fully open source and has accumulated over 25,400 GitHub stars with 35+ contributors and 25,000+ weekly PyPI downloads. For a project that didn't exist before 2024, those numbers are unusual.

More unusual: the team published a peer-reviewed paper on arXiv in January 2025 - "Zep: A Temporal Knowledge Graph Architecture for Agent Memory." It's not a typical startup blog post dressed up in academic formatting. It includes proper benchmarks against alternatives, ablation studies, and honest analysis of tradeoffs. The kind of thing that earns credibility with the engineers doing the evaluations.

The Graphiti MCP Server - which lets Claude Desktop, Cursor, and VS Code Copilot connect directly to Zep's temporal graph - extends the open-source strategy into the developer's daily workflow. When an AI coding assistant remembers the architectural decisions you made last Tuesday and why you rejected the alternative, that's Graphiti working quietly in the background.

This is a classic developer-led growth playbook, but executed with unusual discipline: the open-source project does real work for real developers, which earns the credibility that converts to enterprise evaluations, which becomes the paid product. Chalef ran a version of this with KnowledgeTree in a previous era. He knows how the story goes.

Timeline

From Zero to Infrastructure in 18 Months

What You Can Build

Context Engineering: The Skill That Defines the Next Wave of AI

Zep calls their approach "context engineering" - the discipline of making sure AI agents get exactly the right information at exactly the right time. It's a better frame than "memory," which undersells what's actually happening. Memory implies storing and retrieving. Context engineering implies understanding temporal relationships, inferring what's relevant, and delivering it in the format the LLM needs to reason well.

The practical applications are broad. A sales agent that knows a prospect mentioned budget concerns in the last three calls. A customer support bot that remembers a user's technical setup without asking again. A healthcare agent that tracks how a patient's symptoms evolved over six months. A voice assistant that responds in under 200ms because it doesn't need to process a 50,000-token context window to know what the user just said.

The API is clean and integrates with LangChain, LangGraph, LlamaIndex, Microsoft AutoGen, and Chainlit. SDKs in Python, TypeScript, and Go cover most production environments. The free tier requires no credit card and gives full API access - the developer-friendly entry point that turns evaluators into users.

What Zep is building is foundational: not a feature on top of agents, but part of what makes agents work. In five years, "agent memory" will be a solved category and Graphiti's temporal approach will look obvious in hindsight. The interesting question is who built the standard, and how early they started. Zep started in 2023.