Archive comparison
This comparison belongs to the older memory-layer positioning. Novyx is now focused on governing production-changing agent actions.
Novyx vs Zep
Both provide persistent memory for AI agents. Here's a detailed comparison to help you choose.
TL;DR
Novyx now focuses on governing production-changing agent actions, with memory, Replay, checkpoint recovery, and audit as supporting infrastructure. Zep is a cloud-only service providing conversation memory with classification, user management, and the Graphiti knowledge graph. Zep's Community Edition has been deprecated.
| Feature | Novyx | Zep |
|---|---|---|
| Persistent memory | ✓ | ✓ |
| Semantic search | ✓ | ✓ |
| Checkpoint recovery | ✓ (supported paths) | — |
| Tamper-evident audit trail | ✓ | — |
| Replay (time-travel debugging) | ✓ | — |
| Knowledge graph | ✓ | ✓ (Graphiti) |
| MCP server | Available | — |
| Conversation classification | — | ✓ |
| User/session management | Sessions + Spaces | Built-in |
| Circuit breaker | ✓ (free) | — |
| Python SDK | ✓ | ✓ |
| JS/TS SDK | ✓ | ✓ |
| Free tier | 5,000 memories | Limited free tier |
| Self-hosted | Local SQLite mode | ✗ (CE deprecated) |
| Open source | — | ✗ (CE deprecated) |
Recovery and Audit Context
Zep is built around conversation memory — classifying sessions, extracting facts from dialogue, and managing user context. But it has no mechanism for undoing agent mistakes. When a Zep-powered agent extracts an incorrect fact from a conversation, that bad data persists in your memory store indefinitely.
Novyx keeps recovery as a governed workflow: checkpoint what can be recovered, record the action context, and preserve enough evidence for incident review. That should not be read as a universal transaction over every agent state field and external side effect.
The current Novyx wedge starts before the write: classify the blast radius of a proposed action, block or route approval when risk is high, and keep recovery notes with the decision.
Architecture: Composable API vs Conversation Engine
Zep is designed around conversations. It expects chat-style message arrays, classifies them into categories, and extracts facts automatically. This works well if your agent fits that pattern, but becomes limiting when you need to store structured data, non-conversational observations, or memories from multiple sources.
Novyx is a composable memory API. You store whatever you want — conversations, observations, structured data, relationships — and query it with semantic search. It plugs into LangChain, CrewAI, LlamaIndex, and any custom code. Zep's integrations are more tightly coupled to its conversation-first model.
The practical difference: Novyx adapts to your architecture, while Zep asks you to adapt to its conversation model.
Knowledge Graphs: Built-in vs External
Both platforms offer knowledge graph capabilities. Zep uses Graphiti, a temporal knowledge graph focused on episode-based relationships from conversations. Novyx has a built-in knowledge graph with triple storage (subject-predicate-object) that works with any data source, not just conversations.
Novyx also combines its knowledge graph with rollback — you can revert graph relationships alongside memory state, keeping your entire data layer consistent. Graphiti operates independently from Zep's memory, so rolling back graph changes requires separate manual intervention.
When to Choose Novyx
- ✓You need checkpoint recovery and debugging context for agent state
- ✓You need to govern production-changing actions
- ✓You use Cursor or Claude Code and want MCP integration
- ✓You need audit evidence for incident review
- ✓You want drop-in integrations for LangChain, CrewAI, and LlamaIndex
When to Choose Zep
- ✓You need conversation classification and structured user management
- ✓You want a managed cloud service with no infrastructure to maintain
- ✓You need Graphiti for temporal knowledge graphs
Code Comparison
from novyx import Novyx
nx = Novyx(api_key="YOUR_API_KEY")
# Store
nx.remember("User prefers dark mode", tags=["prefs"])
# Search
results = nx.recall("user preferences")
# Rollback (Novyx-only)
nx.rollback(target="2026-01-15T10:00:00Z")
# Audit trail (Novyx-only)
audit = nx.audit(limit=10)Next protected action
Name the production action your agents cannot perform yet
Start with a 25-minute readiness call for one blocked workflow.