Zep's approach to agent memory is a temporal knowledge graph: conversations and business data combined into entities and relationships that track how things change over time. It's a genuinely interesting architecture — and for some problems, graph structure is exactly right.
The trade-off nobody mentions
Graphs shine when your questions are about relationships over time. But most agent workloads are simpler: "what does the spec say", "what did we decide", "summarize this vendor's docs". For those, a clean knowledge base with hybrid retrieval answers just as well — with far less to set up, tune, and debug.
The Kit for AI approach
- Ingest anything — files, URLs, scans — into knowledge bases as clean Markdown.
- Hybrid retrieval (semantic + keyword) with citations, and chat grounded in your documents.
- remember/recall for session-persistent facts, deduplicated automatically.
- Everything over one API and one MCP server; no graph schema to design.
Which should you pick?
If your product reasons over evolving entity relationships, evaluate a graph seriously — Zep included. If you need agents that reliably read, remember, and cite your content this week, Kit for AI's knowledge-base-first design gets you there with less ceremony. Free to start.